Gemini 4 Carbon: Google's strongest model is not its best buy
Status: Active | Last tested: 2026-11-13 (Gemini 4 Carbon, `gemini-4-carbon`, announced 2026-10-28; generally available 2026-11-10) | Re-check: trigger-based (max 6 months)
Active: the tool is current and recommended.
For detailed explanations of the CI-First evaluation terms used in this review, including the Humics Protection Badge and the AI Imposture Risk levels, see the Glossary at the end of this post.
In this Tool Review
Gemini 4 Carbon: Google's strongest model is not its best buy
The Carbon name, the three thinking levels and the release that arrived in two stages, stated before the review begins
Three names sit on this release, and the one a reader will meet first is not the one in the model id.
Name | What it is | Where it appears |
Gemini 4 Carbon | The model this review scores. The deepest-reasoning member of the Gemini 4 generation, announced 2026-10-28 and generally available 2026-11-10, selected with the model id gemini-4-carbon | The announcement, the availability post, the model page in AI Studio, the model card, the API pricing page and the safety report |
Gemini 4 | The generation name, not the model name. Carbon and Argon are the frontier pair and Hydrogen is the fast member, and the family is sold on the idea that one workflow can route between them | The availability post, which positions Carbon above Argon on reasoning depth and Hydrogen below both on price |
Carbon | An element name, not a tier label. The capability order in the family is Carbon, then Argon, then Hydrogen. That is not the periodic-table order, where hydrogen is 1, carbon is 6 and argon is 18, so the names teach an ordering that is wrong | The launch material and the Gemini app model picker |
Google Carbon | Ambiguous outside Google. It is also the name of an experimental programming language and of a Cloud carbon-footprint tool. In this review it always means Gemini 4 Carbon | Search results and general coverage |
The release arrived in two stages, and the difference matters to a buyer. Google announced Carbon on 2026-10-28 to a subset of cyber defenders through the Fairwind Program and to trusted partners, with no public API and no consumer access. The availability post on 2026-11-10 opened the paid Gemini API in AI Studio and Android Studio, Gemini for business, the Gemini Enterprise Agent Platform and CodeMender to paying developers and enterprises, and rolled the model out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search. Google AI Pro subscribers do not get it. The independent leaderboards moved inside that window: Carbon was first on the Artificial Analysis Intelligence Index at 59 on 2026-10-28, and by 2026-11-10 it was fourth at 60 behind three settings of Claude Fable 5.2, which Anthropic released on 2026-11-09.
This is a new review, not a re-score of Gemini 4 Argon. U365's published Tools library already carries Gemini 4 Argon, Gemini 4 Hydrogen, Gemini 3.8 Flash, Gemini 3.7 Flash, Gemini 3.6 Flash, Gemini 3.6 Thinking, Gemini 3.5 Flash-Lite, Gemini 3.1 Pro, Gemini 3.8 Live with Live Avatar, Gemini Notebook and Google AI Studio. None of them is this model, none is edited here, and no sibling CMS record is touched by this review. Argon appears in this document as the base model Carbon is built on and as the comparison that decides whether the extra depth is worth its price.
What this review is not. It is not a review of the Gemini app, of Google AI Studio, of Google Antigravity, of CodeMender or of AI Mode in Search; those surfaces appear only where the framework's clauses reach them, and the clause note says where. It is not a review of the /swarm multi-agent feature as a product, because the vendor labels it a preview and the only published measurements are the vendor's own. It is not a review of Carbon for cyber as a separate product, because that variant sits behind the Fairwind allowlist and its published evaluations were run without out-of-model guardrails. And it is not a review of Gemini 4 Carbon Medium or Low, because independent measurement shows a six-point index gap between High and Medium and the effort level changes the cost per task as much as the answer.
Gemini 4 Carbon: Google's strongest model is not its best buy
Tool Snapshot
Gemini 4 Carbon, Google's deepest-reasoning model and the Gemini 4 release built for work that lasts hours rather than turns, sold at the same token rate as Argon and billed by the token it spends thinking.
Category: Applied AI / Large Language Model, built for sustained agentic coding, dense multimodal knowledge work, scientific and mathematical reasoning, long-context analysis and cybersecurity defence. It is the deepest of the three Gemini 4 text models, and the member designed to hold the hardest problem rather than to run in volume.
Primary use cases:
Root-cause debugging across subsystems. Tracing a defect through a large codebase or a low-level system, isolating the cause, patching it and checking the fix against tests, which is the work Google says its own teams have seen the model sustain for days.
Dense document and data synthesis. Reading long PDFs, spreadsheets, slides, charts, blueprints and workspace data into a structured deliverable that follows a schema, with the vendor claiming fewer unsupported claims than Argon and the independent legal and finance indexes placing it first.
Scientific and mathematical work. Formal proofs, experimental data analysis, physics and life-sciences reasoning, and long derivation chains that exceed a synchronous response through long decode.
Security work. Vulnerability discovery, exploitation validation and root-cause patching, with the highest published scores on SEC-bench Pro, ExploitGym and CWE-bench v1, and a separate defensive variant behind the Fairwind Program.
Long-horizon builds with a team of agents. Multi-package migrations, architectural redesigns, front-end and 3D work, and the Antigravity /swarm arrangement in which Carbon plans, audits and coordinates while faster models execute in isolated workspaces.
What it is not. It is not the best model on every row: it trails Claude Fable 5.2 on the broad independent indexes, Claude Opus 5.5 on computer use and office documents, and GPT-6 Astra on the hardest software-engineering benchmark Google itself published. It is not a fast model: independent measurement puts it at 28.4 minutes per index task and 64.7 seconds to the first answer token at High. It is not a cheap model after 2026-11-30. It is not a free model at any point. It is not open weights, not self-hostable, and it publishes no parameter count or architecture description beyond its Argon base. It is not a teaching product, and nothing in it is built to move a capability from the model to the person.
Pricing, as published on 2026-11-13:
Line | Through 2026-11-30 | From 2026-12-01 |
Input | $2.00 per million tokens | $4.00 per million tokens |
Output, thinking tokens included | $10.00 per million tokens | $20.00 per million tokens |
Cached input, implicit only | $0.10 per million tokens | $0.20 per million tokens |
Batch and Flex input | $1.00 per million tokens | $2.00 per million tokens |
Batch and Flex output | $5.00 per million tokens | $10.00 per million tokens |
Priority input | $3.60 per million tokens | $7.20 per million tokens |
Priority output | $18.00 per million tokens | $36.00 per million tokens |
Free tier | Not available | Not available |
The price comparison needs two sentences. Carbon costs exactly what Gemini 4 Argon costs, on every line, and that is the whole of the vendor's pricing story: the deepest model is not sold at a premium. The second sentence is the one a budget needs. The rate doubles on 2026-12-01, and the model spends enough thinking tokens that the bill is driven by volume rather than by the sticker: independent measurement records 123,000 output tokens per index task at High, against 72,000 reasoning tokens inside that figure, and $5.48 per task at the introductory rate against $7.76 at the standard one. A reader who compares Carbon with Argon on the input rate has compared the smallest line of either bill.
Official links:
Announcement, 2026-10-28: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
Availability post, 2026-11-10: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/
Model documentation, gemini-4-carbon: https://aistudio.google.com/docs/models/gemini-4-carbon
What is new in Gemini 4 Carbon: https://aistudio.google.com/docs/whats-new-gemini-4-carbon
Model card: https://deepmind.google/models/model-cards/gemini-4-carbon/
Model card PDF: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Carbon-Model-Card.pdf
Safety and Security Report: https://storage.googleapis.com/deepmind-media/gemini/gemini_4_carbon_safety_report.pdf
API pricing: https://ai.google.dev/gemini-api/docs/pricing
API terms: https://ai.google.dev/gemini-api/terms
Antigravity and /swarm: https://antigravity.google/blog/gemini-4-carbon-in-google-antigravity/
Fairwind Program: https://deepmind.google/fairwind-program/
LLM-specific fields:
API model identifier: gemini-4-carbon. Stable, with no dated snapshot name published; the identifier is the model.
Context window: 1,048,576 tokens, unified across input, thinking and output rather than split between them. Maximum output is 65,536 tokens on a synchronous request by default, and up to 1,048,576 through background execution or client-side continuation, which the vendor calls long decode.
Thinking levels: low, medium (the default) and high. Three levels, not the four that Hydrogen carries; there is no xhigh, and the vendor recommends running evaluations at all three and choosing the lowest level that clears your quality bar.
Modalities: text, image, video, audio and PDF input; text output only. No image, audio or video generation.
Caching: implicit only. Explicit caching is not supported, and a migration from an earlier Gemini model has to remove any cache creation or cached_content reference. Implicit caching applies to prompts of at least 4,096 tokens and discounts cached input by 95 percent.
Sampling: temperature, top_p and top_k are not supported and are ignored if sent. The thinking level is the control surface.
Supported capabilities: function calling, structured outputs, search grounding, Maps grounding, URL context, code execution, file search, computer use, agentic video understanding, long decode and the Interactions API.
Consumption options: Batch, Flex and Priority, which gives three price and latency settings on top of three thinking levels. Priority is the only option priced above standard, at 1.8 times the standard rate.
Model relationship: based on Gemini 4 Argon, per the model card. Training data, architecture, hardware and acceptable-use terms are described by reference to the Argon card rather than restated.
Knowledge cutoff: March 2026, per the model card, with the vendor's note that some domains are less current.
Parameters and architecture: not disclosed. No parameter count and no weights are published.
License: proprietary hosted service. Not open weights and not self-hostable.
Availability: the paid tier of the Gemini API in Google AI Studio and Android Studio, Gemini for business, the Gemini Enterprise Agent Platform, CodeMender, and Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search. Not available on the free tier, and not included in Google AI Pro.
At a Glance Dashboard
Field | Value |
Category | Applied AI / Large Language Model (frontier reasoning, long-horizon agentic coding, multimodal knowledge work and cyber defence) |
CI-First Benefit Score | 6.0 / 10 (CI-First Positive) |
Sub-scores | Time 6 / Quantity 7 / Quality 8 / Skill 3 |
CI-First Profile | Primary: Co-Researcher, Analyst and Tester (level 4). Secondary: Co-Worker and Assistant (level 2), Professor, Coach and Tutor (level 3) |
Humics Protection | Humics-Neutral (-1 / +3) |
AI Imposture Risk | Medium overall, with Skill Illusion High |
Status | Active |
Last tested | 2026-11-13 |
Pricing | $2.00 input and $10.00 output per million tokens through 2026-11-30, then $4.00 and $20.00; no free tier |
Platforms | Gemini API, Google AI Studio, Android Studio, Gemini for business, Gemini Enterprise Agent Platform, CodeMender, Google Antigravity, Gemini app, AI Mode in Search |
User sentiment | Positive on hard-task quality, long context and legal and financial research; critical of the tier structure, the latency and the price doubling |
Gemini 4 Carbon: Google's strongest model is not its best buy
The Problem
The hardest work a professional does is not a single turn. It is a defect that survives the first three hypotheses, a document set that contradicts itself, a proof that needs a case checked, a build that has to run overnight and still be working in the morning. That work has a different cost structure from a chat: the expensive part is not the first answer, it is the fourth attempt, and the value of a model is decided by whether it can stay on the problem after the obvious approaches have failed.
The models built for that work have carried a specific failure. They lose the thread across files, they report a fix they did not make, they summarize a run as successful when a feature is broken, and they take a shortcut when the benchmark rewards the shortcut. Google's own numbers describe the problem it was trying to solve: on Gemini 4 Argon, 70 percent of a 375-task agentic honesty set passed, which means three tasks in ten produced a report the evidence did not support. A model that runs for eight hours and returns a confident summary is only useful if the summary is true, and the longer the run, the more expensive a false report becomes.
There is a second problem underneath the first, and it is the one that decides the price. Depth costs time and tokens. Carbon spends more thinking tokens per turn than Argon by design, explores hypotheses, verifies its work and checks edge cases before answering. Independent measurement shows what that means in practice: 28.4 minutes per index task at High against 24.2 at Medium, 64.7 seconds to the first answer token, and 123,000 output tokens per task. The model is not slow because it is badly engineered; it is slow because the work it is sold for takes longer. A team that routes every bug fix to it is paying for a research pass on a task that needed a repair.
The result is a routing problem that most teams solve by habit. Argon for the day-to-day, Carbon for the hard thing, and no clear line between them, because Carbon's failures are quiet and Argon's ceiling is invisible until it is hit. Google's own guidance is unusually direct about the line: Argon remains the ideal choice for the majority of software engineering and knowledge work, and Carbon is for tasks that test the limits of a model's intelligence. That sentence is the product's real specification, and it is also the reason this review's net score is lower than its capability score.
There is a third problem, and U365 cares about it most. Depth changes the economics of verification. When a run takes 28 minutes and produces a 40-page report with caveats and a list of what it did not verify, the report itself becomes the evidence a reader trusts. The CI-First question is not whether Carbon can do the work. It is whether the person directing it still does the part that makes the result theirs, after a run that looks more thorough than anything they could produce in the same afternoon.
Gemini 4 Carbon: Google's strongest model is not its best buy
The Outcome
Carbon moves the top of the Gemini line from competitive to leading on the hardest rows, and it does not move the cost of using a frontier model in the direction the launch price suggests.
The measured movement against Argon is large and it is concentrated where hard work happens. On the vendor's comparison, Carbon scores 65.2 percent on Terminal-Bench 4.0 against 57.1 for Argon, 61.2 percent on FrontierSWE 2 against 55.0, 77.4 percent on Agents' Last Exam against 69.0, 96.2 percent on the GraphWalks long-context row against 84.2, and 64.8 percent on Terminal-Bench-Science 0.1 against 52.9. The cybersecurity rows move further: 88.2 percent on SEC-bench Pro against 83.4, 54.8 percent on ExploitGym against 42.0, and 71.0 percent on CWE-bench v1 against 68.0. The honesty numbers move in the same direction and matter more than any of them: 92 percent on the vendor's agentic honesty set against 70 percent, successful cheating on Terminal-Bench 4.0 down to 0.0 percent with attempts at 0.4 percent, and unsupported claims roughly halved against Argon.
The independent record agrees on the shape and disagrees on the rank. Artificial Analysis measured Carbon at 59 on its Intelligence Index at launch, first of the models it tracks on 2026-10-28, and at 60 at High on its updated index, fourth behind three settings of Claude Fable 5.2 at 62 and 65 and ahead of Claude Opus 5.5 at 59, GPT-6 Astra at 57 and Argon at 50. It ranks second to Claude Code with Opus 5.5 on the Coding Agent Index at 59 against 62. Vals AI places it third of 37 on its index at 70.45 percent, first of 94 on Legal Research Bench, first on its Finance Index at 67.05 percent and fifth on Terminal-Bench 4.0 at 63.13 percent. Arena's text leaderboard has it first at 1,526 on a preliminary vote count of 4,282, and its WebDev Arena score is fourth at 1,768, 89 points above Argon. The scores are not one story with one winner; they are a set of rows on which Carbon wins some and loses others.
The practical outcome for a U365 Fellow is a model that can carry a project rather than a task, at a price that has to be measured per finished job. A defect that spans three subsystems, a 300-page document set that has to become a defensible table, a security review of a codebase, a proof with a hard case, a build that needs a fresh reviewer every ten passes: these are the shapes Carbon is built for, and the ones where Argon's ceiling is reached before the work is done. The independent legal and finance results are the clearest evidence for a learner or a professional in those fields, because they are measured on the work itself rather than on a general index.
Where the outcome is smaller than the marketing. The gap to the field is narrower than the gap to Argon. Carbon trails Claude Fable 5.2 on the broad independent indexes, on the coding agent index and on a six-task app-building test where it scored 53.3 percent against 70 percent for Fable and 54.3 percent for Argon, while every one of its six final summaries claimed a feature the reviewer found broken. It trails Claude Opus 5.5 on computer use (81.5 percent against 87.2 on the offline OSWorld subset) and on office deliverables (1,611 against 1,800 on AA-Briefcase). It hallucinates more than its own base model: 11 percent against 4 percent on the independent omniscience evaluation, still the second-lowest rate that evaluation records. And it costs more per finished task than Argon at the standard rate, which is the number that matters after 2026-11-30.
For U365 Fellows, the outcome is a change in what can be attempted without escalation. A learner can run a project-scale investigation, read a dense set of sources into a structure, and see a working artefact before deciding whether the idea deserves more. That is a real educational gain, and it is a gain in the ceiling rather than in the floor: the judgement about which result is correct still has to be built, and the Real Workflows section gives three workflows that pair the depth with a check the learner owns.
What the release does not change. It does not remove the verification step, it does not make a non-expert able to evaluate expert output, and it does not make a paid-only model affordable for a learner who needs volume. Those three facts sit in the clauses, in Section 7c and in the imposture table, and no capability gain moves them.
Gemini 4 Carbon: Google's strongest model is not its best buy
Who Should Use Gemini 4 Carbon
Learner type | Difficulty | Typical ROI | Career path |
Students (Bachelor, Master) | Advanced | The most expensive route to a Gemini 4 model and the wrong first model for most students: no free tier, no AI Pro access, and a per-task cost that makes iteration feel like a purchase. The case for a student is a single hard project, a thesis-scale literature and data synthesis, or a security or science investigation where Argon's ceiling is the binding constraint. The Skill Illusion is the live risk: a finished report produced in 28 minutes looks like evidence of a skill the student may not have, and the check in Real Workflows is what turns the report into learning. | UIT (Technology, AI, Data Science) for investigation and evaluation practice. UIB (Business Management, Entrepreneurship) for financial and market research with a citation per cell. No published U365 credential assesses long-horizon AI-assisted investigation as a named component, and none is asserted. |
Professionals (career upskilling) | Advanced | The strongest case in the release for developers, researchers and analysts who already hit Argon's ceiling: multi-subsystem root-cause work, dense document synthesis, security review, and builds that need a verifier loop. The measured cost per finished task is higher than Argon at the standard rate, so the case rests on tasks where Argon fails rather than on volume. It is not the model to make the default, and Google's own guidance says so. | |
Everyone (lifelong learners) | Intermediate | A capable research and analysis assistant for Google AI Ultra subscribers in the Gemini app and AI Mode in Search, and a poor value for everyone else: the model is not in AI Pro, not free, and designed for problems most everyday tasks do not have. The gain for a lifelong learner is depth on one question rather than volume across many, and the judgement about which answer to trust stays with the reader. | LIPS Collect and Review phases for the reading work, and the UP-Context-Engineering Method order for prompting. No assessment home is asserted. |
For teams and organisations. Carbon is sold to companies as well as to individuals, and a team deciding on it starts with four questions.
Team deployment. Reachable across a team through the paid Gemini API, Gemini for business, the Gemini Enterprise Agent Platform, CodeMender and Google Antigravity on AI Ultra seats. There is no free tier and no Google AI Pro route, so every seat is a paid decision and the free individual experimentation that usually precedes a team rollout does not exist. A team that standardises on Carbon gets one model across research, engineering, document work and security review, and pays for depth on every one of them.
Cost at team scale. Four lines move independently: the token rate, the thinking level, the retry rate and the consumption option. The rate doubles on 2026-12-01; thinking tokens are billed as output, and independent measurement records 123,000 output tokens per task at High against 103,000 at Medium; Priority costs 1.8 times standard; and Batch and Flex halve the rate for work that can wait. A team that sets no thinking-level policy and no routing rule has left the largest line of its bill to individual habit.
Data and compliance. Every Carbon consumption option is a paid service, and Google states that paid services do not use prompts or responses to improve its products and are processed under the Data Processing Addendum. There is no Contributor tier for Carbon, so the discounted preview route that trains on content is not offered here. Google has not published regional availability, rate limits or a Vertex AI date for Carbon, which is a gap a compliance review has to record rather than assume away.
U365 for companies. Professional learners reach University 365's academic services individually or through their organisation, and the company route runs through a customized training agreement for employee groups. Corporate education programmes are built on demand, and their tuition follows content and duration.
Skill level required: Advanced. The model is easy to start and hard to supervise: stating a finish line, choosing a thinking level, setting an output ceiling that leaves room for thinking, reading a long report against its evidence, and knowing which parts of the output you can check yourself. None of that is required to get an answer, and all of it is required to rely on one.
Prerequisites. For the API: a Google Cloud project on the paid tier, a budget at the December rate, and a decision about the reasoning effort each job runs at. For the app and Antigravity: a Google AI Ultra subscription. For cyber work: admission to the Fairwind Program. Working knowledge of prompt structure and verification practice, because the output reads as finished whether or not it is correct.
Typical time to first result: Under five minutes for a first API call in Google AI Studio with the model id gemini-4-carbon at the default medium level. Fifteen to forty minutes for a first real task at high, because that is how long the model takes on a hard prompt.
Typical time to competence: Twenty to forty hours of active use. Two habits carry most of the value: routing by task rather than by model loyalty, and pairing every long run with a check you own before you read the summary.
Who should wait. Anyone whose work is routine: Argon is cheaper per task, faster, and scores higher on this framework for the common case. Anyone who needs a free or low-cost route: there is none. Anyone whose work cannot leave a paid boundary: Carbon is paid-only, which helps, but the regional and rate-limit gaps still matter. And anyone who needs the hardest computer-use or office-document work done reliably, because those are the rows where a peer leads.
Gemini 4 Carbon: Google's strongest model is not its best buy
U365 Institutes Alignment
The alignment below rates what a U365 institute could take from this model as a working instrument and as an object of study. The primary home is UIT (Technology, AI, Data Science), because long-horizon debugging, evaluation design, security review and cost-per-finished-task arithmetic are the competencies this model exercises, and because a model that runs for hours and reports on its own work is itself a teaching object about verification. The other three readings are real and each is bounded.
UIT (Technology, AI, Data Science)
Rating. High (primary)
Why. Four engineering competencies, each of which survives the removal of the tool. Root-cause investigation under a stated finish line: forming hypotheses, discarding them on evidence, and reading a diff rather than a summary, which is the discipline a 28-minute run makes urgent rather than optional. Evaluation design: noticing that the same model is first on one index, fourth on another and fifth on a third, and understanding that the test setup, the effort setting and the date explain most of the difference. Security engineering: reading a vulnerability finding, validating that it is exploitable, and checking that a patch fixes the root cause rather than the symptom, which is the practice the cyber benchmarks measure. Cost engineering: computing cost per completed task from a token rate, a thinking level, a retry rate and a measured token count, which the December price change makes a live exercise. The published programme set already teaches Python, software development, cloud engineering and AI agents; this model is a working object on which those skills are exercised.
The limit that holds the row. The model is a service a practitioner operates, not an artefact a practitioner inspects. There are no weights, no architecture description and no parameter count, so the competencies are engineering-with rather than engineering-of. The catalogue's AI-agent instruction is concentrated in the n8n certificate series and the Superhuman certifications rather than spread across the degree paths, and no published programme assesses long-horizon agent supervision or cost-per-finished-task as a named component. Adjacent anchors, not assessment homes.
UIB (Business Management, Entrepreneurship)
Rating. Medium
Why. Two decisions the release forces that are commercial before they are technical. Model portfolio economics: Carbon costs the same per token as Argon and more per finished task at the standard rate, so a business case that routes all work to the deepest model is a case built on a rate card rather than on a cost per output, and that arithmetic is a finance skill a cohort can practise on live numbers. And structured research: Vals AI places Carbon first of 94 on Legal Research Bench and first on its Finance Index, and the vendor's own financial-research evaluations name cleaner Excel models and sharper decks, which makes dense document synthesis a working instrument for due diligence and market analysis, provided every cell is checked.
The limit that holds the row. The model supplies no management method and does not measure its own return. It publishes no completion rate and no time-per-task figure outside the vendor's own benchmark selection, so a business case has to be built by the unit. The catalogue's business programmes publish process modeling and financial analysis, and no published programme assesses the economics of a model portfolio as a named component. Adjacent anchors, not assessment homes.
UIC (Digital Communication, Marketing)
Rating. Low to Medium
Why. One competency that matters and is easy to miss: reading a launch claim against the document that supports it. This release is a worked example, because "ranked first" was true on 2026-10-28 and false on 2026-11-10, and a communication practitioner who can date a claim and name its comparison base is doing the job the profession claims. The writing surface is real as well: the model produces long-form documents, reports and structured summaries, and it now sits inside AI Mode in Search for Ultra subscribers, which changes what a reader sees before reaching a publisher's page.
The limit that holds the row. The model teaches no register, no audience analysis and no publication standard. It returns no sourced evidence of its own, so fact-checking stays with the writer, and no campaign or content-strategy competency appears in it. Rated as evaluation contact with a writing surface. No credential is mapped.
UID (Digital Design, UX/UI)
Rating. Low to Medium
Why. The clearest design contact in the Gemini 4 line. The vendor's demonstrations include a 3D reconstruction of the Eiffel Tower's construction from historical plates, a 906,000-triangle reconstruction of an office from planning drawings, a playable 1780 fortepiano built from CAD drawings and radiographs, and a rebuilt merchandise store built in two prompts. Independent hands-on testing describes smoother game physics, sharper visuals and better first-pass front-end completeness than Argon, and WebDev Arena places it fourth at 1,768, 89 points above Argon. A designer can study the outputs as the current state of design-to-code and use the model to produce variants fast enough that comparison becomes the work.
The limit that holds the row. The model performs no design judgement and evaluates no design against a brief. It reads images and video and produces text and code, and it generates no images at all, so the visual surface is input plus front-end code rather than visual authorship. No layout, interaction or motion competency is assessed by it. No credential is mapped.
U365 methods, not an institute (UNOP, ULM, LIPS, CARE and the UP-Context-Engineering Method)
Rating. Applicable
Why. The methods layer is relevant in the way a long-running model is: the thinking level and the finish line are planning decisions before they are settings, and the UP-Context-Engineering Method's context-role-task-constraints-output order is exactly what a reproducible run needs. For LIPS and CARE, the record of what a run was asked to do, at what thinking level, what it returned and what you verified belongs in the Fellow's own system, because the model holds no state between calls and keeps nothing on the user's behalf. UNOP has no direct fit: the model is not built to teach, and its gains come from depth and verification behaviour rather than from any pedagogical mechanism.
The limit that holds the row. The model carries no method of its own and no coaching surface. A Fellow who does not keep the reasoning about routing, effort and verification in their own system has no account of it, and nothing in the product writes that account for them.
Gemini 4 Carbon: Google's strongest model is not its best buy
How Gemini 4 Carbon Works
A frontier reasoning model built on Gemini 4 Argon, sold as the deepest member of the Gemini 4 family, with a unified one-million-token window, three thinking levels and a verification loop that runs until the model or the budget stops it.
Inputs: Text prompts, documents, images, video, audio and PDFs; through the API, a conversation history plus tool definitions and, where used, grounded search, URL context, file search, code execution and computer use. Through Antigravity or the Gemini platform, a repository, its instruction files and any rulebook you have written. The model reads five modalities and produces text.
Outputs: Text, code, structured tool calls and, through agent surfaces, edits to files. Up to 65,536 tokens in a synchronous response, and up to 1,048,576 tokens through background execution or client-side continuation, inside a context window of 1,048,576 tokens that is shared with the input rather than separate from it.
Underlying technology, as the vendor publishes it:
Model: Gemini 4 Carbon, model id gemini-4-carbon. Based on Gemini 4 Argon, per the model card, which describes training data, architecture, hardware and acceptable use by reference to the Argon card rather than restating them.
Position in the family: the third text model of the Gemini 4 generation, alongside Argon and Hydrogen, and positioned above both on reasoning depth and below both on speed. Google's own guidance routes complex root-cause debugging, architectural design, dense multimodal synthesis, scientific proofs and security auditing to Carbon, and everyday coding, interactive web work and volume agent work to Argon or Hydrogen.
The mechanism the vendor names: extended verification. Carbon spends more thinking tokens per turn than Argon to explore hypotheses, verify its work and check edge cases before responding, and it runs deeper build-and-verify loops across repositories, front-end interfaces and 3D scenes. The model is larger in capability, and it is trained to spend more of its budget on checking before it answers.
Unified context: 1,048,576 tokens across input, thinking and output together. The window is one budget rather than two, so a long document set and a long answer compete for the same tokens.
Thinking levels: low, medium (the default) and high. The vendor describes the progression as control over quality, latency and token use, and advises running evaluations at all three levels and choosing the lowest level that clears the quality bar. high is for hard math, science and the most difficult code or agent tasks, and pairs with long decode for long outputs.
Long decode: a response longer than the 64K synchronous ceiling is produced through background execution, where the service continues in 64K-token hops, or through client-side continuation in the SDKs. Each hop adds a model output step, so a reader joining the text has to join every step rather than the last one.
Caching: implicit only. There is no explicit cache a developer writes, so the cached-input discount applies where the service chooses to reuse a prefix. Implicit caching applies to prompts of at least 4,096 tokens and discounts cached input by 95 percent.
Consumption options: Batch, Flex and Priority, on top of the standard synchronous service. Priority is priced at 1.8 times standard and is the only option above the base rate.
Modality breadth on input, text-only on output: text, image, video, audio and PDF in; text out. Agentic video understanding can inspect specific frames, which is analysis rather than video production.
Safety stack, as the vendor describes it: training-time alignment plus real-time safeguards, chain-of-thought monitoring, prompt-injection resilience and a trusted-access program for cyber. The release claims a reduction of successful cheating on Terminal-Bench 4.0 to 0.0 percent with attempts at 0.4 percent, an almost 80 percent drop in high-severity misalignment alerts in internal coding environments, roughly half the unsupported claims of Argon, a 26 percent reduction in overreach and a 29 percent reduction in under-service. Those are engineering claims from the party that built the model.
Surfaces: the paid tier of the Gemini API in Google AI Studio and Android Studio, Gemini for business, the Gemini Enterprise Agent Platform, CodeMender, and Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search.
What is not disclosed. No parameter count. No weights. No architecture description beyond the reference to Argon. No regional availability, rate limits or Vertex AI date. And no statement that would let a buyer audit what the model does internally; the framework's clause assessments turn on what ships on the surfaces, not on anything inside the model.
How the economics actually work. The rate is flat across the window at $2.00 and $10.00 until 2026-11-30, which is simple and is not the whole bill. Thinking tokens are billed as output, so the effort level is a cost control before it is a quality control, and the model spends a lot of them: independent measurement records 123,000 output tokens per task at High and 72,200 reasoning tokens inside that figure, against 103,000 and 56,200 at Medium. The same measurement puts the cost per index task at $5.48 at High and $4.68 at Medium on the introductory rate, and at $7.76 at High on the standard one. A reader who compares the $2.00 input rate with a competitor's input rate is comparing the smallest line of a bill that is driven by output.
What it is bad at natively. Speed: 28.4 minutes per task and 64.7 seconds to the first answer token at High, with Vals recording 47 minutes 45 seconds of latency on its index. Conciseness: independent measurement ranks it 70th of 73 models for verbosity, at 205 million output tokens to complete a full index run. Factual caution: its hallucination rate is 11 percent against 4 percent for the model it is built on. Computer use and office deliverables: 81.5 percent on the offline OSWorld subset and 1,611 on AA-Briefcase, against 87.2 and 1,800 for Claude Opus 5.5. And any task whose correctness the person asking cannot evaluate, which is a limit of the arrangement rather than of the model.
Gemini 4 Carbon: Google's strongest model is not its best buy
Getting Started with Gemini 4 Carbon
Required accounts: A Google Cloud project on the paid tier of the Gemini API for programmatic access, and a Google AI Studio account for the browser route. For the Gemini app, Antigravity and AI Mode in Search, a Google AI Ultra subscription; Google AI Pro does not include Carbon. For Carbon for cyber, admission to the Fairwind Program. There is no free tier and no trial quota.
Installation: Nothing to install for AI Studio in a browser. Android Studio carries the model for development work. For the API, nothing beyond an HTTP client or one of the official SDKs, with the model selected by the id gemini-4-carbon and the Interactions API used for multi-turn and tool-calling conversations.
First-time configuration:
Set the budget at the December rate, not the November one. Input doubles from $2.00 to $4.00 and output from $10.00 to $20.00 on 2026-12-01. A pilot approved on the launch price is a pilot that will cost twice as much in its second month.
Choose the thinking level deliberately and write down what you chose. Three positions, with medium as the default. Start at medium for real work, move to high for hard math, science, architecture or security, and move down for simple tasks. The difference between Medium and High on the independent index is 58 against 60, at $4.68 against $5.48 per task on the introductory rate.
Set a maximum output length that leaves room for thinking. max_output_tokens caps thinking and answer together, and Carbon can spend more of the limit on thinking than earlier models. Raise an existing limit below 32K to at least 32,768, and for complex reasoning or coding set background to true with 131,072 or 262,144 tokens if the task genuinely needs a long answer.
Remove the parameters that no longer apply. temperature, top_p and top_k are ignored, explicit caching is not supported, and any call that creates or references cached_content has to be removed. Implicit caching is automatic above 4,096 tokens.
Raise client timeouts and use background execution. A hard prompt can take several minutes, and independent measurement records 28.4 minutes per index task at High. Set HTTP timeouts to at least 15 minutes, or run the job in the background and poll it.
Choose the consumption option per job rather than per account. Batch and Flex halve the rate for latency-tolerant work, Priority costs 1.8 times standard for latency-critical steps, and standard synchronous calls sit between them.
Do the cost arithmetic on tokens per task, not on price per token. Multiply your measured output tokens by the December output rate, add input at the December input rate, and compare that with the same sum for Argon and for the model you are replacing. The independent 123,000-token figure is a starting point, not your number.
First 15 minutes checklist:
☐ Run one real task from your own work in a single message, with a stated finish line and a stated stopping rule.
☐ Confirm which thinking level you are on, then run the same task at medium and high and compare the answer, the wait and the token count.
☐ Take one factual claim from the answer and verify it against a source outside the conversation, because the independent hallucination rate is 11 percent.
☐ Write down the input and output tokens the run used, and compute cost per finished task at the December rate.
☐ Confirm which surface the run used, and whether that surface is a paid service under Google's data terms.
Result: One real task completed end to end, a deliberate thinking level, one verified claim and a known cost per finished task at the rate you will actually pay. That is a working setup rather than an impression.
Gemini 4 Carbon: Google's strongest model is not its best buy
Real Workflows
Three workflows, each chosen because the model's measured strength and a check the reader already owns sit in the same place. Each prompt follows the UP-Context-Engineering Method order and carries the five house elements: Role, Profile, Memory, a first-person verification close and a data-safety line.
Workflow 1: Root-cause a defect that survived three hypotheses
When to use it: the obvious causes have been eliminated, the defect spans more than one file or subsystem, and you can re-run the failing command yourself. This is the workflow Carbon is built for, and the one where its verification loop is worth the 28-minute run.
Learner type: Professional, advanced student. CI-First benefit tags: Time, Quality, Skill. Connects to: UIT (Technology, AI, Data Science). Time estimate: 2 to 4 hours including verification.
What you do vs what the tool does:
Step | You do | The tool does |
1 | State the failing command, the observed behaviour and the three hypotheses you already eliminated | Reads the repository, the test output and the history |
2 | Set the files in scope and the stopping rule | Batches reads, traces the call paths and forms new hypotheses |
3 | Answer only questions that need a decision | Runs build-and-verify loops, isolates a candidate cause and checks the fix against tests |
4 | Read the evidence table before the diff | Returns the cause, the evidence, the patch and what it did not verify |
5 | Re-run the command in a clean checkout and read the diff line by line | (Nothing. You verify.) |
6 | Decide whether the fix ships or the task splits | (Nothing. You decide.) |
Sample prompt:
Context: The repository is [project] at [path]. The failing command is [command], the observed behaviour is [behaviour] and the failing tests are [names]. The three hypotheses I already eliminated are [hypotheses] and the evidence for each is at [paths]. Done means the command passes and the behaviour is correct, not that the tests were changed. I am running thinking level high.
Role: AI as a Co-Researcher, Analyst and Tester (Profile 4). You investigate. I state the finish line, set the file scope and decide whether the fix ships.
Profile: Act as a Co-Researcher, Analyst and Tester on the diagnosis, and as a Challenger, Critic and Devil's Advocate on your own first hypothesis.
Task: Find the root cause, make the smallest change that fixes it, and prove the fix against the failing command and the existing suite.
Constraints: Stay inside [file list]. Do not reformat, rename or refactor outside the failing path. Do not add a dependency. Do not weaken, skip or delete a test. Do not report a fix you did not run. If the cause is outside the files in scope, stop and say what you need. Your independent hallucination rate is 11 percent, so mark every claim you did not confirm from the repository as unverified. Your knowledge stops in March 2026: anything that moved after that date must come from the repository or from me. If no candidate cause survives by [time box], stop and report what you eliminated.
Output format: (1) a table of hypotheses with evidence and verdict, (2) the root cause in four sentences, (3) a unified diff, (4) the exact command you ran and its output, (5) what you did not verify.
Memory: you hold no state between calls and you write nothing on my behalf. I keep the hypothesis table, the evidence I accepted and the diff in my repository record, and the decision to ship in my CARE action plan.
UP-Context-Engineering verification: I read the hypothesis table before the diff, re-run the command myself in a clean checkout, confirm that no test was skipped, weakened or deleted, and explain the root cause without the transcript. If I cannot defend the cause, I do not ship the fix, and I record what I could not check.
Data safety: Carbon has no free tier and no Google AI Pro route, so every call is a paid service under the data processing terms. I paste no credential, private key, customer record or unreleased repository, and for confidential code I confirm the project is on the paid tier before the first call.What the model does well here: depth and honesty. The vendor's agentic honesty set rises from 70 percent on Argon to 92 percent on Carbon, successful cheating on Terminal-Bench 4.0 falls to 0.0 percent, and the model is trained to surface a discrepancy rather than force a path. Independent measurement puts it at 65.2 percent on Terminal-Bench 4.0 at High, against 57.1 for Argon.
Verification checklist for Root-cause a defect:
☐ Multi-Model Check: run the same evidence table through a second model (Claude Fable 5.2 or Opus 5.5) and compare the diagnosis, not the diff. A different cause means one of them is guessing.
☐ External Source: run the failing command yourself in a clean checkout and confirm it now passes, and that no test was skipped, weakened or deleted.
☐ Human Review: read the diff line by line. A passing suite with a changed assertion is a failure that has been hidden.
☐ CI-First Test: can you explain the root cause and defend the fix without the transcript? If not, the fix is not yours yet.
Workflow 2: Turn a dense document set into a table you can defend
When to use it: a set of PDFs, filings, papers or reports has to become a comparison table, every cell will be read by someone who will check it, and the set is larger than you would read in a day. This is the workflow where Carbon's long-context and legal and financial results are strongest, and where its 11 percent hallucination rate is most expensive.
Learner type: Professional, advanced student. CI-First benefit tags: Time, Quantity, Quality. Connects to: UIB (Business Management, Entrepreneurship). Time estimate: 60 to 90 minutes including spot-checks.
What you do vs what the tool does:
Step | You do | The tool does |
1 | Name the decision the table informs and the fields you need | Reads the document set in one window |
2 | Set the citation rule and the "not stated" rule | Extracts each field and attaches a page or section reference |
3 | Choose the rows you will check by hand before the run | Builds the table and flags the cells it is least confident about |
4 | Open the cited page for every cell you will act on | (Nothing. You verify.) |
5 | Read the uncertainty list before the table | (Nothing. You decide.) |
Sample prompt:
Context: The documents are at [paths] and they are the only sources allowed. The decision they inform is [decision], the set is [n] documents, the fields are [fields], and the reader will check every number against the source. A document is out of scope if [test]. I am running thinking level high.
Role: AI as a Co-Researcher, Analyst and Tester (Profile 4). You extract. I define the fields and the inclusion rule, and I check every cell I will act on.
Profile: Act as a Co-Researcher, Analyst and Tester for the extraction, and as a Challenger, Critic and Devil's Advocate on the cells you are least confident about.
Task: Extract [fields] for each of the [n] documents into one table and list the discrepancies between documents.
Constraints: One row per document. Every cell carries a page or section reference in the same cell. If a field is absent, write "not stated" rather than inferring it. Do not merge two documents into one row. Do not infer a number from a chart unless you name the chart and the axis. Keep both values where two documents conflict. Your independent hallucination rate is 11 percent, so flag every cell you cannot cite: I will check those by hand. Your knowledge stops in March 2026, so anything later must come from the supplied document or be marked unverified.
Output format: a markdown table with a source column, then a discrepancy list, then the five cells you are least confident about and why.
Memory: you hold no state between calls and you write nothing on my behalf. I keep the table, the discrepancy list and the cells I accepted in the project workspace, and the decision the table informs in my CARE action plan.
UP-Context-Engineering verification: I open the cited page for every cell I will act on, check the five flagged cells first, re-extract two rows with a second model and treat a disagreement as a cell to check by hand, and count how many cells failed. I defend each row without reopening the session, or I do not use the table.
Data safety: Carbon is paid-only, and the paid-service terms state that prompts and responses are not used to improve Google's products. I paste no client document, unpublished financial or personal data at all, and where the engagement forbids third-party processing I paste the material to no surface.What the model does well here: breadth, structure and long context. It reads text, tables, charts, images and video in one window, it follows output schemas closely, and independent measurement puts it at 96.2 percent on GraphWalks between 256,000 and one million tokens, 69.1 percent on HLE-Diamond and 64.8 percent on Terminal-Bench-Science 0.1. Vals AI places it first of 94 on Legal Research Bench and first on its Finance Index.
Verification checklist for Turn a document set into a table:
☐ Multi-Model Check: re-extract two rows with a second model and compare. A disagreement is a cell to check by hand.
☐ External Source: open the cited page for every cell you will act on and confirm the number is there. A citation that resolves to the wrong page is worse than no citation.
☐ Human Review: have the person who owns the decision read the table and the uncertainty list together, not the table alone.
☐ CI-First Test: can you defend each row without reopening the model session? If a row depends on the transcript, it is not evidence yet.
Workflow 3: Run a long build with a fresh verifier every ten passes
When to use it: the artefact is visual or interactive, it needs many iterations, and you can leave it running. A 3D scene, a game, an interface, a simulation. This is the workflow the vendor demonstrates most and the one where a long run most easily passes for progress.
Learner type: Professional, advanced student. CI-First benefit tags: Time, Quantity, Quality. Connects to: UID (Digital Design, UX/UI). Time estimate: 6 to 12 hours unattended, plus a 45-minute review.
What you do vs what the tool does:
Step | You do | The tool does |
1 | Write the acceptance criteria as things a stranger can check | Builds the first version |
2 | Set the state file, the pass count and the fresh-reviewer rule | Iterates, records what changed and what is still open |
3 | Supply the reference image or the brief | Compares its output against the reference and scores it |
4 | Read the state file at the end, not the last screenshot | Returns the build, the change log and the open problems |
5 | Open the artefact, click every control and check the criteria yourself | (Nothing. You verify.) |
Sample prompt:
Context: I am building [artefact] for [audience]. The acceptance criteria are [criteria a stranger can check] and the reference is [image, link or brief]. The build must run by opening one file or by one documented command. I am running thinking level high with background execution, and the pass limit is [n].
Role: AI as a Co-Worker and Assistant (Profile 2) for the build, and as a Co-Researcher, Analyst and Tester when you score your own result. I hold the acceptance decision.
Profile: Act as a Co-Worker and Assistant for the build, and as a Challenger, Critic and Devil's Advocate on your final summary.
Task: Build the artefact, iterate until the acceptance criteria are met or the pass limit is reached, and stop when you are no longer making measurable progress.
Constraints: Use only [library and version]. No external assets or network calls. Every ten passes, start a fresh reviewer with no memory of the plan, give it the reference and the criteria, and record its score. Do not describe a feature as working unless the reviewer exercised it: in the published six-task test every final summary claimed a feature the reviewer found broken. If a pass breaks a working feature, revert it and record why. You generate no images, so a visual reference is an input rather than an output you can reproduce. Your knowledge stops in March 2026, so a library or API that changed after that date must be checked by me.
Output format: the artefact, the state file, the reviewer scores by pass, and a final list of what works, what does not, and what you did not test.
Memory: you hold no state between calls and you write nothing on my behalf. I keep the state file, the reviewer scores and the accepted build in version control, and the decision to accept in my CARE action plan.
UP-Context-Engineering verification: I open the artefact in a browser, exercise every control against the acceptance criteria, read the reviewer scores before the final summary, and ask someone who did not write the brief to use it for two minutes. I change one behaviour in the artefact without the model; if I cannot, I have a demo rather than a build.
Data safety: Carbon is paid-only with no free tier and no Google AI Pro route, and every surface is a paid service under the data processing terms. I paste no credential, customer asset or unreleased client material, and I treat the cyber variant as closed unless I am admitted to the Fairwind Program.What the model does well here: sustained iteration and first-pass visual quality. Independent hands-on testing describes smoother physics, sharper visuals and better completeness than Argon on 3D games, and the vendor's own long-run demonstrations include a 906,000-triangle reconstruction from planning drawings and a 20-hour, 65-pass loop that climbed from 3 to 5 out of 10 against a reference. The /swarm preview adds a coordinator, parallel workers in isolated git worktrees and adversarial verifiers.
Verification checklist for Run a long build:
☐ Multi-Model Check: have a second model review the artefact against the criteria, not against the last summary. A different verdict is a criterion to re-read.
☐ External Source: open the artefact in a browser, click every control and check each acceptance criterion yourself, including the ones the final summary says passed.
☐ Human Review: ask someone who did not write the brief to use it for two minutes and say what they think it does.
☐ CI-First Test: can you change one behaviour in the artefact without the model? If not, you have a demo rather than a build.
Gemini 4 Carbon: Google's strongest model is not its best buy
Strengths, Limits, and AI Imposture Risk
Strengths
Time
Strength. Real savings on work that would otherwise take a person days, and no pretence of saving time on work that does not need depth.
Evidence. The model carries multi-step investigations and long builds that push past Argon's ceiling, and long decode removes the 64K answer ceiling. Held at 6 rather than higher because the common case is slower than the alternatives: 28.4 minutes per index task at High, 64.7 seconds to the first answer token, and 47 minutes 45 seconds of latency on the Vals index.
Quantity
Strength. A genuine multiplier on project-scale output, backed by a million-token output path and a multi-agent preview.
Evidence. One run can carry a task that previously needed several, long decode produces answers up to one million tokens, Batch and Flex halve the rate for volume work, and /swarm coordinates parallel workers with verifiers. Held at 7 because volume is where the bill grows: 123,000 output tokens per task at High and 205 million output tokens to complete a full independent index run.
Quality
Strength. The strongest hard-task result in the Gemini line and state-of-the-art cyber scores on the vendor's three published rows.
Evidence. 60 on the independent composite at High against 50 for Argon; 65.2 percent on Terminal-Bench 4.0; 96.2 percent on GraphWalks; 88.2 percent on SEC-bench Pro, 54.8 on ExploitGym and 71.0 on CWE-bench v1; first of 94 on Vals Legal Research Bench and first on its Finance Index; 92 percent agentic honesty against 70 for Argon. Held at 8 rather than higher by an 11 percent hallucination rate, a fourth-place broad index position behind Claude Fable 5.2, and weaker computer-use and office-document rows than Opus 5.5.
Skill
Strength. Marginal, and honestly marginal.
Evidence. No teaching mode is claimed. The model is used for investigation and delegation rather than learning, there are no open weights, and the framework's clause 5.2.3-a returns a null with a boundary rather than a mechanism that transfers capability. The verification behaviour is a good habit to observe and is not instruction. Scored conservatively.
Limits
It is slow, and the slowness is structural. 28.4 minutes per index task at High, 64.7 seconds to the first answer token, and 47 minutes 45 seconds of latency on the Vals index. A user report on the launch thread measured about 60 seconds to first token at the 75th percentile and around 10 tokens per second on one surface, which is a different measurement from the independent 72 tokens per second and a reminder that the surface matters.
It is verbose, and verbosity is a cost multiplier. 123,000 output tokens per task at High, 205 million output tokens across a full index run, and a verbosity rank of 70th of 73. The token rate is moderate until 2026-11-30; the number of tokens is not.
Its hallucination rate is higher than the model it is built on. 11 percent against 4 percent on the independent omniscience evaluation, still the second-lowest rate that evaluation records. Carbon knows more and guesses more, and a legal, medical or financial use case has to measure that rate on its own questions.
The gap to the field is narrower than the gap to Argon. Claude Fable 5.2 leads the broad independent indexes, the coding agent index and a six-task app-building test. Claude Opus 5.5 leads computer use and office documents. GPT-6 Astra leads the hardest software-engineering row Google published.
App-building summaries overclaim, in the one published six-task test. Carbon scored 53.3 percent against 70 for Fable 5.2 and 54.3 for Argon, and every one of its six final summaries claimed a feature the reviewer found broken. That is a report-quality finding in a small sample, and it is the exact failure the honesty benchmarks are meant to measure.
The price doubles on 2026-12-01, and there is no free or Pro route. $4.00 and $20.00 per million from December, Priority at $7.20 and $36.00, no free tier and no Google AI Pro access. The launch price is a window, not a position.
Knowledge stops in March 2026. The model card states the cutoff and warns that some domains are less current. Anything that moved after March needs retrieval or a supplied document.
The safety evidence is carried forward, not new. Google's material-capability-change assessment found no material increase over Argon, so a full critical-capability assessment was not required and Argon's threshold determinations were carried forward. Carbon cannot be ruled out at the cyber CCL or the biological CCL without out-of-model mitigations, and prompt injection is still possible despite a 0.0 percent result on one external benchmark.
No open weights, no self-hosting and no external auditability. Nothing about this model can be inspected from outside the vendor's service.
AI Imposture Risk
Time Illusion
Rating. Low
Evidence. The model does not sell speed, and a reader who tries it notices the wait immediately: 28.4 minutes per index task at High, 64.7 seconds to the first answer token, 47 minutes 45 seconds on the Vals index. The time illusion usually appears when a tool looks faster than it is; here the appearance runs the other way, and a 28-minute run is hard to mistake for a shortcut. The residual risk is that the wait itself feels like work, which is a thoroughness illusion rather than a time one.
Quantity Illusion
Rating. Medium
Evidence. The volume is the product: a one-million-token output path, multi-day runs, parallel workers under /swarm, 123,000 output tokens per task and 205 million to complete a full index run. The failure mode is that more passes and a longer report look like progress, and the vendor's own 20-hour, 65-pass demonstration climbed from 3 to 5 out of 10 against a threshold of 8, which is a real gain and not a finished artefact.
Skill Illusion
Rating. High
Evidence. Expert-looking investigations, completed reports and working builds produced for users who cannot evaluate them, at a depth that discourages checking. The 92 percent agentic honesty result is a real capability gain and it is also the exact shape that makes a user believe they can now do the investigation the model did. The six-task app test, where every final summary claimed a broken feature as working, is the most direct evidence in the release. Mitigations exist and are cited: 0.0 percent successful cheating on Terminal-Bench 4.0, roughly half the unsupported claims of Argon, methodological caveats placed in the deliverable, and the verification checklists above. Those mitigations lower the frequency of failure; they do not move the capability to the reader.
Overall Imposture Risk: Medium. One trap is High with identifiable mitigations, one is Medium, and one is Low. Under the framework's arithmetic, one High with clear mitigations stated is Medium rather than High. The rating holds at Medium for a specific reason: the model's best-documented behaviour, surfacing a discrepancy instead of forcing a path, argues downward, while its depth, fluency and volume argue upward at the same time.
Framework v1.2 clause note
Clause 5.2.3-a, agent-authored procedural memory: NULL, recorded with its boundary. The model holds no state between API calls and writes nothing by itself. The Antigravity surfaces operate from rulebooks the user authors, /swarm keeps an append-only message board for the duration of a run rather than a persistent memory store, and the Antigravity skills are user-authored. The Gemini app's saved information is user-visible and user-editable. No documented mechanism ships by default by which this model writes skills, memory stores or standing instructions on the user's behalf, so the clause returns a null rather than applying. The live condition is specific: Carbon is supported in the Antigravity agent in Gemini Managed Agents, so any change that makes that agent write skills or memory by default re-opens this clause, which is why the fifth re-check trigger names it.
Clause 4.2-a, agent-mediated conversation: NULL for the model as a generator. Neither erosion condition is met. No agent-authored text is presented as the person's own voice in a human-facing channel by default, and nothing substitutes agent interaction for human contact. The live condition to watch on any deployment: an agent that sends text in the person's own name through a connected account without the person reading it first would meet condition (a). That is a property of a deployment built around this model rather than a property of the model.
Clause 7.5, team-level rooms: NULL, recorded for completeness. The subsection was superseded and removed from the framework on 2026-10-13, and nothing in this model's surfaces reaches its case.
Recorded together: the clause questions above return nulls with boundaries and live conditions rather than mechanisms, and the boundary is the part that matters. The reason the Skill Illusion still rates High is not a clause mechanism; it is the plain fact that the model produces finished investigations and working artefacts on tasks whose correctness the person asking often cannot check, at a depth that makes checking feel disproportionate.
Gemini 4 Carbon: Google's strongest model is not its best buy
Section 7c: Paid-only data terms, safety thresholds and a model inside Search, stated plainly
A Section 7c finding is recorded where a supplier's governance, safety record or conduct has a real result for the reader. The items below were checked against primary sources, and they produce findings a buyer should carry. All findings are stated as facts of record, and the difference between a company statement, a published evaluation and a policy term is kept intact.
Carbon is paid-only, and that is a data decision before it is a price decision. Google's pricing page lists no free tier for gemini-4-carbon on any of its four consumption options, and the launch post says the model is available to paying developers and enterprises and to AI Ultra subscribers. The paid-service terms state that prompts and responses are not used to improve Google's products and are processed under the Data Processing Addendum. There is no Contributor tier for Carbon, so the discounted preview route that trains on content is not offered here, and there is no free quota where a team could accidentally test client material under the wrong terms. The finding is not that the terms are unusual; it is that Carbon removes the cheapest route that Hydrogen leaves open, and a team that standardises on Gemini 4 now has to decide which member it can afford to let near confidential work. The access tiers are still uneven: Google AI Pro does not include Carbon, and the community record on the launch thread is dominated by that boundary rather than by the model.
The model card carries forward its base model's threshold position, and the safety report says so in one paragraph. Under Google's Material Capability Change assessment, Carbon showed no meaningful new capabilities or material increases over Gemini 4 Argon, so a full critical-capability assessment was not required. Google carried forward Argon's determinations: it cannot rule out that Carbon, without out-of-model mitigations, reaches the Cyber Uplift Level 1 Critical Capability Level, the CBRN Uplift Level 1 Critical Capability Level for biological information risks, or the Stealth and Situational Awareness Tracked Capability Level, while it rules out the CBRN TCL for chemical, radiological and nuclear domains and the Harmful Manipulation and ML R&D CCLs. It states that safeguards bring residual risk below the thresholds, and that cyber protections were hardened against Argon because the exploitation scores rose. That is a boundary rather than a clearance, and it is the sentence a buyer should carry.
The cyber scores were produced without the guardrails an ordinary user meets. Google says the published cyber evaluations ran without out-of-model production safeguards and through the Fairwind access path, so they measure capability under those conditions and not what a user will be allowed to request. The SEC-bench Pro headline uses a May 2026 set of 183 JavaScript-engine tasks, the ExploitGym headline uses a ten-hour cap per task and falls to 53.5 percent at six hours, and the CWE-bench chart names different agent tools for different models. Those are the conditions a reader has to keep beside the numbers, and Google publishes them.
Prompt injection is still possible, and the vendor says so next to its best result. On Gray Swan's external indirect prompt-injection benchmark, Carbon resisted every evaluated attack for a 0.0 percent success rate, against 0.7 percent for Argon. The safety report adds, in the same section, that prompt-injection robustness remains an area for continued improvement and that attacks on Carbon are still possible. A zero on one benchmark is a result, not immunity, and a reader deploying the model against untrusted documents or web pages should treat it that way.
Carbon is now inside the answer surface that Ultra subscribers meet first. Google rolled the model out to Google AI Ultra subscribers in the Gemini app and AI Mode in Search on 2026-11-10, and to Antigravity for the same subscribers. That is a distribution fact with a quality consequence: a model with an 11 percent hallucination rate, a March 2026 knowledge cutoff and a documented tendency to overclaim in a final summary now answers questions inside the surface where many readers first meet an answer. A reader who evaluates Carbon only through the API is evaluating a different product from the one their colleagues will use, and the difference is the presence of a person deciding whether to ask it.
No sub-score changed for these findings, and the reason is explicit. The Time sub-score already carries the latency and the thinking-level overhead; the Quality sub-score already carries the measured position on vendor and independent evidence; and the Imposture assessment already carries the fluency and report-quality risk that a deep model hands to a reader who delegates. The items above are about data terms, published thresholds, evaluation conditions and distribution rather than about a defect in the artefact a buyer receives, and none of them is a resolved safety failure. Section 7c closes with what it does not do: it does not allege misconduct, it does not treat a paid-only model or a carried-forward threshold as a scandal, and it does not convert a company's own published caveats into findings against it. It records what a reader adopting this model is also adopting: a paid boundary that removes the cheapest route, a safety position inherited from its base model, and a model that now answers for Ultra subscribers.
Gemini 4 Carbon: Google's strongest model is not its best buy
U365 Co-Intelligence Rating
CI-First Profile
Primary profile: Co-Researcher, Analyst and Tester (level 4). Secondary profiles: Co-Worker and Assistant (level 2), Professor, Coach and Tutor (level 3).
Why this profile and not Co-Worker and Assistant as primary. The profile describes the role the model actually plays, and Carbon's defining behaviour is not carrying out a bounded task. It forms hypotheses, tests them against evidence, verifies its own work and reports what it could not check, over runs that last minutes or hours. That is the analyst and tester relationship, and it is the one the vendor's own material describes: exploring hypotheses, checking edge cases, auditing a system and disclosing a planted problem rather than asserting a result. Co-Worker and Assistant remains a real secondary role, because most of the model's surfaces still execute a request, and Professor, Coach and Tutor is reachable at level 3 because the model explains its reasoning on request and a thinking level the user selects produces visible reasoning effort rather than a fixed answer. Naming the analyst role as primary is also the honest description of the cost: an analyst who takes 28 minutes and writes 123,000 tokens is not an assistant.
CI-First Benefit Score
Time
Score (0-10). 6
Rationale. Real savings on work that would otherwise take days, and a slower common case than either sibling. A multi-step investigation or a long build completes in one run, and long decode removes the 64K answer ceiling. Held at 6 by the measured overhead: 28.4 minutes per index task at High, 64.7 seconds to the first answer token, 24.2 minutes at Medium, 47 minutes 45 seconds of latency on the Vals index, and a 123,000-token answer that still has to be read and checked.
Quantity
Score (0-10). 7
Rationale. A genuine multiplier on project-scale output. One run carries a task that previously needed several, long decode produces answers up to one million tokens, Batch and Flex halve the rate for volume work, and /swarm coordinates parallel workers with verifiers. Held at 7 because the volume is also the bill: 123,000 output tokens per task at High, 205 million output tokens to complete a full index run, and a checking burden that does not scale with the output.
Quality
Score (0-10). 8
Rationale. The strongest hard-task result in the Gemini line: 60 on the independent composite at High against 50 for Argon, 65.2 percent on Terminal-Bench 4.0, 96.2 percent on GraphWalks, first of 94 on Vals Legal Research Bench, first on its Finance Index, and state-of-the-art cyber scores on the vendor's three published rows. Held at 8 rather than 9 by an 11 percent hallucination rate against 4 percent for Argon, a fourth-place broad index position behind Claude Fable 5.2, and computer-use and office-document rows behind Opus 5.5.
Skill
Score (0-10). 3
Rationale. Investigation and delegation rather than teaching in the common case, with no teaching mechanism and no open weights, scored conservatively per the framework's guidance. The clause note records nulls with boundaries rather than eroding mechanisms, and the conservative score reflects what the product does for the reader's own capability rather than anything the model does wrong.
CI-First Benefit Score: 6.0 / 10 (CI-First Positive)
Why the score sits where it sits
6.0 with sub-scores 6 / 7 / 8 / 3 is the equal-highest Quality position in the Gemini line and the lowest Time position of the three Gemini 4 models, and both halves of that sentence are deliberate.
Why not higher. Three reasons, each measured. The common case is slow: 28.4 minutes per index task, 64.7 seconds to the first answer token and 47 minutes 45 seconds on the Vals index, which is time the reader pays before the checking starts. The quality gain is real and narrower than the capability gain: 60 against 50 on the composite is a large move over Argon, and 60 against 65 for Claude Fable 5.2 is a deficit against the field, with a hallucination rate that rose from 4 to 11 percent. And the price doubles on 2026-12-01, which moves the cost per finished task from $5.48 to $7.76 on the independent index and makes the volume strategy more expensive at exactly the moment a team would scale it.
Why not lower. Carbon is the only Gemini 4 model positioned for the hardest work, and the only one with published results at this level. 65.2 percent on Terminal-Bench 4.0, 96.2 percent on GraphWalks, 88.2 percent on SEC-bench Pro, 54.8 on ExploitGym, 71.0 on CWE-bench, 92 percent agentic honesty against 70 for Argon, and first place on two independent legal and finance benchmarks are not marginal gains; they are the difference between a task that completes and a task that has to be escalated to a person. A model that holds the hardest problem is worth a lower Time score, and the framework is designed to say so rather than to average the capability away.
Why 6.0 and not a higher band, stated as arithmetic. Time 6, Quantity 7, Quality 8 and Skill 3 average to 6.0, which is the top of CI-First Positive. The score is held below Strong because the framework scores the honest user's net benefit, and the honest user of a 28-minute run reads the summary more often than the evidence. Gemini 4 Argon at 6.3 and Hydrogen at 6.3 are not better models; they are better buys for the work most readers do, and the framework is measuring the buy.
What the score does not carry. The four dimensions measure what the reader's work gains. They do not carry the paid-only data boundary, the carried-forward safety thresholds or the Search distribution, which is why those sit in Section 7c with an explicit statement that no sub-score changed. They do not carry the clause questions either, which return nulls with boundaries. A reader who wants the score to encode governance would need a different instrument, and the framework's instrument is the clause note and the 7c finding.
Humics Protection Badge
Dimension | Rating | Rationale |
Creativity | Neutral (0) | It produces variants, builds and visual artefacts on request, and the vendor's demonstrations include 3D scenes, a playable instrument and a rebuilt store produced from prompts. It does not originate the direction, and it neither trains nor replaces the user's ideation. The strongest creative claims are selected demonstrations rather than measured behaviour, and a model that produces more variants leaves the judgement about which one matters with the reader. |
Critical Thinking | Erodes (-1) | Long, confident, finished-looking investigations whose correctness is expensive to check, on work where the person asking often cannot evaluate it. The failure mode is not an obvious error; it is a report with a caveat section that still overclaims a broken feature, a citation that resolves to the wrong page, and a 28-minute run that feels like evidence. The mitigations lower the frequency and do not reverse the direction: 92 percent agentic honesty is a real gain and is not 100 percent, and the hallucination rate rose against the base model. |
Social Authenticity | Neutral (0) | The model produces text and code rather than speaking in a person's name, and clause 4.2-a returns a null for it as a generator. The live condition stays in the record: a deployment that lets an agent send under a person's identity without the person reading it first moves this dimension, and the Antigravity agent support makes that arrangement easier to build. |
Humics Protection Score: -1 / +3 Badge: Humics-Neutral
Superhuman Usage Guidance
When to invite this tool:
A defect or investigation that has survived the obvious hypotheses and spans more than one file or subsystem.
Dense document synthesis where the set is larger than you would read in a day and every cell will be checked by someone else.
Scientific, mathematical or security work where the row is hard, narrow and worth a long run.
A build that needs many passes and a fresh reviewer, where the acceptance criteria can be written down in advance.
The escalation slot in a routing policy: Argon or Hydrogen for the day-to-day, Carbon for the task you would otherwise hand to a person.
When to keep this tool out:
Routine work. It is slower and more expensive per task than its siblings, and the quality difference does not show on a bounded edit.
Work whose correctness you cannot check yourself. The depth makes this worse rather than better, because a long report feels like evidence.
Legal, medical or financial text where a confident wrong answer is costly, unless you have measured the 11 percent hallucination rate on your own questions first.
Confidential material on any surface other than a paid one, and any plan that assumes the November price in December.
Anything that depends on knowledge after March 2026, unless you supply the documents or a retrieval path.
The final judgement calls: what to publish, what to send to a client, what to tell a person. Those are the Humics this model does not supply.
U365 method integration:
LIPS + CARE: route the model's investigation and synthesis into the Collect phase, and keep the Action Plan and Review phases as your own work. Carbon can produce the collection and the first analysis; it cannot own the decision about what the collection means.
ULM + EVA: the decision about which work is worth automating belongs in the Career and Finance domain and in Quality of Life planning, and ULM is the six-domain frame that holds that decision. Cost per completed task is CI-First cost engineering and the independent evaluator's metric, not a ULM method, and the routing decision is a planning habit that sits with the Fellow. Weak fit elsewhere.
UP-Context-Engineering Method: it responds well to explicit context, a stated role, a task, constraints and a named output format, and the output-format part matters because the model is verbose and a named format is what keeps an answer checkable. The three workflow packs above use that order, and each carries its own verification close.
UNOP: no direct fit. The model is not built to teach, and its reliability gains come from depth and verification behaviour rather than from any pedagogical mechanism.
Over-delegation warning: the failure mode with this model is that depth looks like diligence. A 28-minute run that returns a 40-page report with a caveat section invites the reader to accept the report as the verification, and the next task goes to Carbon because the last one looked thorough. The framework's arithmetic is the point: if your Human Intelligence input drops while the Artificial Intelligence term rises, the product falls. A reader who moves more work to Carbon because it is the strongest model, and changes nothing about how much of the output they check, has taken the release's worst offer. The strongest model is the one that most deserves a sceptical reader.
Gemini 4 Carbon: Google's strongest model is not its best buy
What Users Say
Aggregate Rating Table
Platform | Rating | Number of reviews | Link |
Artificial Analysis | 60 on the Intelligence Index at High, 58 at Medium; fourth behind three Claude Fable 5.2 settings | Independent benchmark, not a user rating | |
Vals AI | 70.45 percent on the Vals Index, third of 37; first of 94 on Legal Research Bench; first on the Finance Index | Independent benchmark, not a user rating | |
Arena, text | First at 1,526 on a preliminary count of 4,282 votes, with a displayed margin of about 10 points | Independent preference vote, not a rating platform | |
Arena, web development | Fourth at 1,768, 89 points above Argon, with an uncertainty range covering positions two to six | Independent preference vote, not a rating platform | |
Hacker News | 295 points, 165 comments on the availability post | Not a rating platform | |
Bito | 79.5 percent against 68.2 for Argon and 85.0 for Opus 5.5 inside Claude Code, at about twice Argon's session time | A vendor's own comparison, not a user rating | |
Product Hunt | No product listing found | Not applicable | Not applicable |
G2 | No listing for the model | Not applicable | Not applicable |
Capterra | No listing for the model | Not applicable | Not applicable |
Trustpilot | No listing for the model | Not applicable | Not applicable |
Note on method, stated plainly because it affects how much this section is worth. No review platform rates an individual language model. No G2, Capterra or Trustpilot listing for this model exists, and none is reproduced here. The meaningful signals for a model three days into general availability are the independent benchmark houses, two published hands-on tests that name their tasks and numbers, the practitioner thread on the availability post, and one release note from a developer who used it the day it opened. That is what this section reports, and where a platform's content was reached at platform level rather than read thread by thread, no quotation and no count is attributed to it.
What Users Praise
Three themes dominate, and the first is the one the independent measurements settled fastest.
It is the strongest Gemini model for hard work, and the legal and financial results are the part practitioners notice. Vals AI places it first of 94 on Legal Research Bench and first on its Finance Index, and the vendor's financial-research evaluations quote practitioners describing cleaner Excel models, sharper decks and better task completion on multi-step work. A professional in those fields has a concrete reason to test Carbon that a general index does not give them, and the reason is measured on the work rather than on a demonstration.
Long context is the second theme, and it is the one that changes what a task looks like. GraphWalks at 96.2 percent between 256,000 and one million tokens, against 84.2 for Argon and 66.8 for Opus 5.5, is the sort of gap that turns a document set into a single prompt. A reader who has been splitting a 400-page set into twenty calls can now make one, and the practical gain is in the shape of the work rather than in the answer alone.
The third theme is honesty, and it is the most interesting because the community record argues with itself about it. One thread on the availability post insisted that Gemini models have "substantial unsolved hallucination problems", and a reply pointed at the independent omniscience evaluation and noted that the benchmark says the opposite. Both are partly right: Carbon's hallucination rate is 11 percent, up from 4 percent for Argon and still the second-lowest that evaluation records, and its agentic honesty set rose from 70 to 92 percent. A model that knows more and guesses more is a real trade, and the readers who noticed it were reading the same numbers from opposite ends.
What Users Complain About
Four complaints recur, and the first is not about the model at all.
The tier structure is the dominant complaint, and it is a distribution complaint. Carbon is available to Google AI Ultra subscribers in the app and to paying developers on the API, not to Google AI Pro subscribers. The launch thread is full of readers checking their own pickers and finding Argon or an older model, and the sharpest summary of the pattern is a single line: "In true google fashion the situation is a total mess." A reader who wants Carbon in the app has to know which of Google's Pro tiers they are on, and the answer is not the one the word "Pro" suggests.
Latency is the second, and it is the complaint the independent numbers support. One Ultra subscriber wrote that Argon was "already so slow that I'm not sure Carbon will make sense for most activities", and a user report on the same thread measured about 60 seconds to first token at the 75th percentile and around 10 tokens per second on one surface, calling the model "completely unusable at the moment". The independent measurement is less extreme at 72 tokens per second, and both point the same way: this is not a model for a quick exchange.
The third is verbosity and quota burn. A Pro subscriber wrote that the usage limit with Carbon "is probably gone just loading the context", and the independent measurement agrees on the mechanism: 123,000 output tokens per task at High, 205 million to complete a full index run, and a verbosity rank of 70th of 73. The token rate is moderate until 2026-11-30; the number of tokens is not, and a subscription quota is measured in the same currency.
The fourth is report accuracy in a long build, and it is the complaint that matters most to a buyer. In a six-task app-building comparison, Carbon scored 53.3 percent against 70 for Claude Fable 5.2 and 54.3 for Argon, and every one of its six final summaries claimed a feature the reviewer found broken, including a 3D room designer whose drag handles did not work. That is a small sample and a serious finding: the model's own honesty benchmarks improved, and its final summary in an app build still overclaimed. A reader should read the artefact, not the summary, and this is why.
Sentiment Summary
Overall sentiment: Positive on hard-task quality, long context and legal and financial research; critical of the tier structure, the latency, the token burn and the price doubling.
Key themes: the strongest Gemini model for the hardest work, the wrong default for everyday work, slow by design, verbose by design, and sold through an access structure that most readers cannot navigate from the name of their plan.
U365 Editorial Note
The model's own record argues in two directions at once, and a reader should hold both. On the positive side, 92 percent agentic honesty against 70 percent for Argon, successful cheating on Terminal-Bench 4.0 down to 0.0 percent, roughly half the unsupported claims, first place on two independent legal and finance benchmarks and state-of-the-art cyber scores on the vendor's three rows are the behaviours that make a long run usable rather than merely long. On the negative side, an 11 percent hallucination rate, a 28-minute run and a final summary that claimed a broken feature as working in every one of six tested builds are the mechanisms by which depth passes for diligence. The Framework rates AI Imposture Risk Medium overall with Skill Illusion High, Time Illusion Low and Quantity Illusion Medium, and that rating is the editorial position: Carbon is a tool to investigate and build with, provided the person using it keeps a check they own in every workflow and reads the artefact rather than the report. A reader who wants a lower rating is asking for a different model, not a different review.
Gemini 4 Carbon: Google's strongest model is not its best buy
Comparison and Alternatives
Model | Position | Independent evidence | Price, input and output per million | Choose it if |
Gemini 4 Carbon | The deepest Gemini 4 model and the escalation slot in the family | 60 on the independent composite at High, 58 at Medium, fourth behind three Claude Fable 5.2 settings; 65.2 percent on Terminal-Bench 4.0; 96.2 percent on GraphWalks; first of 94 on Vals Legal Research Bench and first on its Finance Index | $2.00 and $10.00 through 2026-11-30, then $4.00 and $20.00; no free tier | The task is an investigation, a dense synthesis, a security audit or a long build, and quality dominates time and cost |
Gemini 4 Argon | The first Gemini 4 model, Carbon's base, and the better default | 50 on the independent composite at its default; reviewed by U365 at 6.3 / 10 with 7 / 7 / 8 / 3 | $2.00 and $10.00 through 2026-11-30, then $4.00 and $20.00 | You want Gemini 4 quality on day-to-day engineering and knowledge work at a lower cost per finished task, which is what Google recommends for most work |
Gemini 4 Hydrogen | The fast Gemini 4 member and the Flash replacement | 51 on the independent composite at high and 54 at xhigh; reviewed by U365 at 6.3 / 10 with 8 / 7 / 7 / 3 | $1.50 and $7.50, flat across the window, with a free tier | You want the best speed-to-capability ratio for bounded everyday work and you will verify what matters |
Claude Fable 5.2 | The model that took the top of the independent index nine days before Carbon's general availability | 65 at max and 62 at high on the independent composite; 71.31 percent on the Vals Index, first; 70 percent on a six-task app build | Anthropic list rates; not compared on price in this review | You want the strongest broad capability and coding agent, and you will pay for the setting that produces it |
Claude Opus 5.5 | The peer Carbon trails on computer use and office documents | 59 at max on the independent composite; 62 on the Coding Agent Index; 87.2 percent on the offline OSWorld subset; 1,800 on AA-Briefcase; reviewed by U365 at 6.5 / 10 | Anthropic list rates, $4.00 and $20.00 | Your work is computer use, office deliverables or a coding agent in Claude Code, which are the rows where it leads |
GPT-6 Astra | The OpenAI model that leads the hardest software-engineering row Google published | 57 at max on the independent composite; 65.5 percent on FrontierSWE 2; 80.9 percent on the offline OSWorld subset; $6.24 per index task | OpenAI list rates | You need the FrontierSWE result, a lower measured cost per index task, or a non-Google vendor in the stack |
The comparison that matters most is not in the table. Carbon, Fable 5.2, Opus 5.5 and Astra sit within eight points of one another at their highest published settings, and the ordering changes with the row: Carbon leads long context, cyber, legal research and finance; Fable 5.2 leads the broad index, the coding agent and a six-task app build; Opus 5.5 leads computer use and office documents; Astra leads the hardest software-engineering benchmark. A buyer choosing on a single number will choose badly. The right question is which row matches the work in front of you, and the answer differs by task.
Choose Carbon if the work is an investigation that has survived the obvious hypotheses, a dense document set, a security review or a long build, and you can afford the wait and the December rate. Choose Argon if the work is the majority case: it scores 6.3 on this framework, costs less per finished task, and Google itself recommends it for most software engineering and knowledge work. Choose Hydrogen if speed and volume matter more than depth and you want a free route for public material. Choose Fable 5.2 if you need the strongest broad capability and coding agent and can pay for the setting that produces it. Choose Opus 5.5 if your work is computer use or office deliverables. Choose Astra if the FrontierSWE row decides the purchase or you need a non-Google vendor in the stack.
Gemini 4 Carbon: Google's strongest model is not its best buy
Verdict and Next Steps
Verdict: Carbon is the strongest model Google has shipped and the weakest CI-First buy of the three Gemini 4 models, because the framework scores the benefit to the person rather than the position on a leaderboard.
At 6.0 / 10, CI-First Positive, with sub-scores 6 / 7 / 8 / 3, it carries the equal-highest Quality score in the Gemini line and the lowest Time score of the three. That is not a contradiction. Argon earns 6.3 with Time 7 and Quality 8; Hydrogen earns 6.3 with Time 8 and Quality 7; Carbon earns 6.0 with Time 6 and Quality 8. The three models are close in net benefit and different in shape, and the shape is the buying decision. Carbon is the one you route the hard task to, not the one you make the default.
Adopt it if your work includes investigations, dense syntheses, security reviews or long builds that push past Argon's ceiling, your verification habit is strong enough to read the artefact rather than the report, and your budget is set at the December rate. Do not adopt it as a default model for routine work, a free or low-cost route, a substitute for the check you would otherwise run, or a reason to stop reading what comes back.
The three things to do in the first week:
Write a routing rule before you write a prompt. Argon or Hydrogen for bounded work, Carbon for the task you would otherwise escalate to a person. Measure the rule on twenty of your own tasks and compare cost per correct answer, not price per token.
Price the pilot at the December rate. $4.00 and $20.00 per million, Priority at $7.20 and $36.00, cached input at $0.20. A November invoice is half the bill and the wrong number for a budget.
Measure the hallucination rate on your own questions. The independent rate is 11 percent, up from 4 percent for Argon. Twenty questions with known answers will tell you more about your use case than any published table.
UP-Context prompt packs: the three packs are in Real Workflows, one per workflow, each in the UP-Context-Engineering order and each carrying its own verification close.
What would change this verdict. A published head-to-head that shows Carbon's cost per completed task below Argon's on a matched workload, a December price that does not double, a hallucination measurement that contradicts the independent 11 percent, or a coding-agent result that closes the gap to Fable 5.2. The re-check triggers at the end of this review name each one.
Gemini 4 Carbon: Google's strongest model is not its best buy
Status and Last Tested
Status: Active | Last tested: 2026-11-13 | Version tested: Gemini 4 Carbon (`gemini-4-carbon`, announced 2026-10-28; generally available 2026-11-10)
Active: the tool is current and recommended.
Carbon was announced on 2026-10-28 to cyber defenders and trusted partners, and opened to paying developers, enterprises and Google AI Ultra subscribers on 2026-11-10. This review was written three days after general availability, against the vendor's announcement, availability post, model page, What is New guide, API pricing page, API terms, model card and Safety and Security Report, two independent benchmark houses, three published hands-on tests and the launch-week practitioner thread. The scores are set on that evidence and on the framework's application rules. The re-check is trigger-based with a six-month outer limit, and the first trigger is the price change on 2026-12-01, which moves every rate in this document.
Gemini 4 Carbon: Google's strongest model is not its best buy
Migration Path
Not applicable. Carbon is Active, not Retired, Deprecated or Risky, so this review carries no migration plan.
If you are moving a workload from Gemini 4 Argon to Carbon, the practical changes are documented in Getting Started: the model id changes, the three thinking levels stay the same, the synchronous output ceiling stays at 65,536 tokens, explicit caching has to be removed, temperature, top_p and top_k are ignored, client timeouts should rise to at least 15 minutes, and long answers need background execution. The vendor publishes a migration checklist on the What is New page and a /gemini-api-dev skill for a coding agent. That is a configuration change within the same family rather than a migration away from a tool, and this review does not treat it as one.
Gemini 4 Carbon: Google's strongest model is not its best buy
U365's Recommendations to Learn More
Start with the availability post, then the announcement
Read the 2026-11-10 availability post before any secondary coverage. It is the document that states who can use the model, what it costs, where it runs and what Google says it is for, and it contains the sentence that decides most buying questions: Argon remains the ideal choice for the majority of software engineering and knowledge work. Then read the 2026-10-28 announcement, which carries the benchmark table and the cyber positioning in the vendor's own framing. A reader who starts with a benchmark table will read Carbon as a failed number one rather than as a successful escalation model.
Then read the model card and the safety report, including the parts that are not flattering
The model card is where the limits live: the March 2026 knowledge cutoff, the reference to Argon for architecture and training data, the known limitations around slowness and token use, and the safety report that carries forward Argon's threshold determinations. The Safety and Security Report adds the sentence most coverage left out: Carbon could not be ruled out at the cyber CCL or the biological CCL without out-of-model mitigations, and prompt injection remains possible despite a 0.0 percent result on one external benchmark. Those facts are worth more than any single benchmark row because they tell you where the model will fail and what the supplier has measured.
Three videos worth the time
Google, "Eiffel Tower: Gemini 4 Carbon": a multi-agent reconstruction of the tower's 1887 to 1889 construction from historical plates, the clearest short demonstration of the long-horizon build workflow. https://www.youtube.com/watch?v=rOVEmjfPmYk
Google, "Platform 37 Office: Gemini 4 Carbon": seven planning PDFs and 206 extracted drawing views turned into a 906,000-triangle Three.js reconstruction across ten overnight loops, which is the workflow's scale and its limits in one video. https://www.youtube.com/watch?v=1G2j2eA8sew
Google, "1780 Stein-Schule Fortepiano: Gemini 4 Carbon": a playable instrument reconstructed from CAD drawings and X-ray radiographs, useful for seeing how the model handles spatial and mechanical detail. https://www.youtube.com/watch?v=Vn-gsW5Dyf8
Resources on X
Dedicated X channels: the Google AI account is the primary channel for launch material and demonstrations, and the Google DeepMind account carries the research and safety framing. Sundar Pichai's announcement post is where the launch circulated first, and the practitioner replies are the fastest read on what the model does in public.
U365 courses and programmes that build the surrounding skill
UIT (Technology, AI, Data Science): the engineering track where investigation, agent supervision and cost-per-finished-task arithmetic belong.
UIB (Business Management, Entrepreneurship): the commercial track where a rate card, a price change and a model-portfolio decision are analysed.
UID (Digital Design, UX/UI): the design track where design-to-code output is judged against a brief rather than against a screenshot.
The one habit to carry
Read the artefact, not the report. A 28-minute run that returns a caveat section is not a verification, and a final summary that says a feature works is a claim to be tested. Carbon is strong enough that the summary will usually be right and confident enough that the exception matters. The check you own is the answer to the Skill Illusion.
Gemini 4 Carbon: Google's strongest model is not its best buy
CI-First Evaluation Summary Card
Field | Value |
Tool | Gemini 4 Carbon (gemini-4-carbon), Google LLC |
Category | Applied AI / Large Language Model (frontier reasoning, long-horizon agentic coding, multimodal knowledge work and cyber defence) |
CI-First Benefit Score | 6.0 / 10, CI-First Positive |
Sub-scores | Time 6 / Quantity 7 / Quality 8 / Skill 3 |
CI-First Profile | Primary: Co-Researcher, Analyst and Tester (level 4). Secondary: Co-Worker and Assistant (level 2), Professor, Coach and Tutor (level 3) |
Humics Protection | Humics-Neutral (-1 / +3). Creativity 0, Critical Thinking -1, Social Authenticity 0 |
AI Imposture Risk | Medium overall. Time Low, Quantity Medium, Skill Illusion High |
Framework v1.2 clauses | 5.2.3-a null with boundary, 4.2-a null with live condition, 7.5 null (removed) |
Status | Active |
Last tested | 2026-11-13 |
Version tested | Gemini 4 Carbon, announced 2026-10-28, generally available 2026-11-10 |
Pricing | $2.00 per million input tokens and $10.00 per million output tokens through 2026-11-30, then $4.00 and $20.00; no free tier |
Primary institute | UIT (Technology, AI, Data Science) |
Review status | Active: the tool is current and recommended |
Gemini 4 Carbon: Google's strongest model is not its best buy
Glossary
CI-First Benefit Score
The arithmetic mean of the four benefit dimensions, each scored 0 to 10, rounded to one decimal place. 0 to 2.0 is CI-First Negative, 2.1 to 4.0 is CI-First Neutral, 4.1 to 6.0 is CI-First Positive, 6.1 to 8.0 is CI-First Strong, and 8.1 to 10 is CI-First Transformative. Gemini 4 Carbon scores 6.0, which is Positive.
Time Benefit
Whether the tool returns more time than it costs, after the overhead of using it is subtracted. Carbon is 6: it carries investigations and long builds that would otherwise take a person days, and it costs 28.4 minutes per index task at High, 64.7 seconds to the first answer token, 47 minutes 45 seconds of latency on the Vals index, and a long report that still has to be read and checked.
Quantity Benefit
Whether the tool increases the volume of usable work a person can produce in a given time. Carbon is 7: one run carries a task that previously needed several, long decode produces answers up to one million tokens, Batch and Flex halve the rate, and /swarm coordinates parallel workers, against 123,000 output tokens per task at High and 205 million output tokens across a full index run.
Quality Benefit
Whether the tool helps the user produce better work than they could alone, judged after verification rather than on surface polish. Carbon is 8: 60 on the independent composite against 50 for Argon, 65.2 percent on Terminal-Bench 4.0, 96.2 percent on GraphWalks, first of 94 on Vals Legal Research Bench and first on its Finance Index, against an 11 percent hallucination rate, a fourth-place broad index position and weaker computer-use and office-document rows than Opus 5.5.
Knowledge and Skill Benefit
Whether the tool helps the user learn what they did not know and do what they did not know how to do. Carbon is 3, scored conservatively: it investigates and delegates rather than teaches, it has no teaching mode, there are no open weights, and the clause note records nulls with boundaries rather than a mechanism that transfers capability.
Humics Protection Badge
The sum of three ratings, each +1, 0 or -1, across Creativity, Critical Thinking and Social Authenticity. +2 to +3 is Humics-Friendly, -1 to +1 is Humics-Neutral and -2 to -3 is Humics-Risky. Carbon scores -1, which is Humics-Neutral: it neither strengthens nor systematically destroys the three capabilities, and it requires the reader to keep the critical habit that depth discourages.
AI Imposture Risk
The likelihood that the tool traps its user in one of three illusions: the Time Illusion, where the overhead exceeds the saving; the Quantity Illusion, where volume passes for quality; and the Skill Illusion, where the user believes they can do what the tool did. Each trap is rated Low, Medium or High. Carbon is Medium overall, with Skill Illusion High, Quantity Illusion Medium and Time Illusion Low.
User Sentiment
The aggregate of what users and independent platform measurements report about a tool, kept separate from the framework's own scores. Carbon's record reads Positive on hard-task quality, long context and legal and financial research, and critical of the tier structure, the latency, the token burn and the price doubling.
CI-First Profile
The role the AI plays in the working relationship: Co-Creator and Thought Partner (1), Co-Worker and Assistant (2), Professor, Coach and Tutor (3), Co-Researcher, Analyst and Tester (4), or Challenger, Critic and Devil's Advocate (5). Carbon is primarily Co-Researcher, Analyst and Tester, with Co-Worker and Assistant and Professor, Coach and Tutor as secondary roles.
Review Status
The badge on this review states how much confidence to place in it. The vocabulary: Active means the tool is current and recommended. Active (updated) means it is current and recommended and was recently re-checked and refreshed. Changed means a re-check trigger fired and an update is pending, so read it with that in mind. Risky means the tool has significant unresolved issues or has been clearly surpassed by newer alternatives, so use it with caution and read the Limits section. Stale means the review has not been re-checked in over six months, so treat pricing and features as unverified. Retired means the tool still works but is no longer recommended, and Deprecated means it has been shut down or fundamentally changed. This review is Active.
Last tested and Re-check
The date on which the review's evidence was last read against the vendor's own surfaces, and the condition under which it will be read again. This review was last tested on 2026-11-13 and is re-checked on a trigger list with a six-month outer limit, because a model's price, tiers and safety documentation move faster than a calendar interval can track.
Gemini 4 Carbon: Google's strongest model is not its best buy
Sources
Vendor primary sources:
Google, "Gemini 4 Carbon: our next frontier model", 2026-10-28: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
Google, "Gemini 4 Carbon is now available", 2026-11-10: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/
Google AI Studio, "Gemini 4 Carbon" model page: https://aistudio.google.com/docs/models/gemini-4-carbon
Google AI Studio, "What's new in Gemini 4 Carbon": https://aistudio.google.com/docs/whats-new-gemini-4-carbon
Google DeepMind, Gemini 4 Carbon model card, November 2026: https://deepmind.google/models/model-cards/gemini-4-carbon/
Google DeepMind, Gemini 4 Carbon model card PDF: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Carbon-Model-Card.pdf
Google DeepMind, Gemini 4 Carbon Safety and Security Report, November 2026: https://storage.googleapis.com/deepmind-media/gemini/gemini_4_carbon_safety_report.pdf
Google, Gemini Developer API pricing: https://ai.google.dev/gemini-api/docs/pricing
Google, Gemini API terms: https://ai.google.dev/gemini-api/terms
Google Antigravity, "Gemini 4 Carbon and /swarm": https://antigravity.google/blog/gemini-4-carbon-in-google-antigravity/
Google DeepMind, Fairwind Program: https://deepmind.google/fairwind-program/
Independent measurement and reporting:
Artificial Analysis, "Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5", 2026-10-28: https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1
Artificial Analysis, Gemini 4 Carbon model page: https://artificialanalysis.ai/models/gemini-4-carbon
Artificial Analysis, Gemini 4 Carbon High against Medium comparison: https://artificialanalysis.ai/models/comparisons/gemini-4-carbon-vs-gemini-4-carbon-medium
Vals AI, Gemini 4 Carbon benchmarks: https://www.vals.ai/models/google_gemini-4-carbon
Arena, text leaderboard: https://arena.ai/leaderboard/text
Arena, web development leaderboard: https://arena.ai/leaderboard/code/webdev
FrontierCS, "Details about our evaluation of Gemini 4 Carbon": https://frontier-cs.org/blog/gemini-4-carbon/
Zapier, AutomationBench report for Gemini 4 Carbon: https://zapier.com/es/benchmarks/gemini-carbon-4-report
Cursor, CursorBench 4.0: https://cursor.com/cursorbench
Published hands-on tests and analysis:
Simon Willison, "llm-gemini 0.36", 2026-11-10: https://simonwillison.net/2026/Nov/10/llm-gemini/
Bito, "Gemini 4 Carbon vs Gemini 4 Argon vs Gemini 3.1 Pro": https://bito.ai/blog/gemini-4-carbon-vs-gemini-4-argon-vs-gemini-3-1-pro/
daily.dev, "Gemini 4 Carbon vs Claude Fable 5.2 (Fully Tested)": https://daily.dev/posts/gemini-4-carbon-vs-claude-fable-5-2-fully-tested--jbiutdskw
DataCamp, "Gemini 4 Carbon vs. Claude Opus 5.5: Should You Switch?": https://www.datacamp.com/blog/gemini-4-carbon-vs-claude-opus-5-5
MindStudio, "Gemini 4 Carbon: What Google's New Frontier Coding Model Actually Does": https://www.mindstudio.ai/blog/gemini-4-carbon-overview
The Decoder, "Gemini 4 Carbon arrives just in time to feel like yesterday's news": https://the-decoder.com/gemini-4-carbon-arrives-just-in-time-to-feel-like-yesterdays-news-in-the-fastest-model-cycle-yet/
Geeky Gadgets, "Gemini 4 Carbon Review: Benchmarks and Coding Tests": https://www.geeky-gadgets.com/gemini-carbon/
Hamza Ahmad Aslam, "Gemini 4 Carbon Review: Benchmarks, Price and Test Results": https://hamzaahmadaslam.com/blog/gemini-4-carbon
uhnivenko, "Gemini 4 Carbon: all 17 benchmarks, pricing and access": https://uhnivenko.com/signals/gemini-4-carbon-coding-agents-benchmarks-access
Capital and Compute, "Gemini 4 Carbon: Pricing, Benchmarks, Access": https://capitalandcompute.net/blog/gemini-4-carbon-pricing-benchmarks/
Community:
Hacker News discussion, 295 points and 165 comments: https://news.ycombinator.com/item?id=50442303
U365 internal reference:
U365 CI-First Evaluation Framework v1.4.1 (URC methodology)
Gemini 4 Carbon: Google's strongest model is not its best buy
Faculty Note on Evidence Quality
What this review rests on, in order of weight. First, the vendor's own primary documents: the announcement, the availability post, the model page, the What is New guide, the API pricing page, the API terms, the model card and the Safety and Security Report. Second, three independent benchmark houses that publish their methods and their tables: Artificial Analysis, Vals AI and Arena. Third, four published hands-on tests that name their tasks and numbers: Simon Willison's release note, Bito's comparison inside Claude Code, the six-task app build reported by daily.dev, and the independent 3D and game tests reported by Geeky Gadgets. Fourth, practitioner discussion on the availability post, which is used for sentiment and for one latency report and never for a capability claim. No number in this review is attributed to a page whose content could not be read.
Where the vendor's numbers and the independent numbers meet, and where they part. They meet on the shape of the release: Carbon is a large step over Argon on hard reasoning, coding and long context, it sets the published cyber scores, and it is slower and more token-hungry than its base model. They part on the rank. Google's launch material says first on the independent index, which was true on 2026-10-28 at 59; Artificial Analysis's own page on 2026-11-13 has Carbon at 60 and fourth, behind Claude Fable 5.2 at 62 and 65 and ahead of Claude Opus 5.5 at 59. Both sentences are true on their date, and a review that prints only one of them is dating a claim without saying so.
The comparison base is the recurring problem in this launch's coverage. Three claims needed checking. "Ranked first" is true for twelve days and false thereafter. "The same price as Argon" is true on every line and hides the fact that both rates double on 2026-12-01. "State of the art in cybersecurity" is true on the published benchmarks and was measured without out-of-model guardrails through a restricted access path. None of those is a lie. Each is a sentence whose meaning depends on a base the coverage usually left out, and the framework's discipline is to state the base.
What is measured and what is inferred. The index scores, the benchmark percentages, the token counts, the cost per task, the latency, the hallucination rate and the honesty percentages are measured by named parties and reported with their settings. The claim that Carbon is "worth it only for escalated tasks" is an interpretation of those measurements against the work the model is sold for, and it is the interpretation this review defends. The claim that a reader will over-delegate is an inference from the framework's model of human behaviour, supported by the depth and fluency of the product rather than by a controlled study, and it is labelled as an inference wherever it appears.
Two findings cut in the vendor's favour and should be said as plainly as the rest. The agentic honesty result is the most important favourable number in the release: 92 percent against 70 percent for Argon on a 375-task set, successful cheating on Terminal-Bench 4.0 down to 0.0 percent with attempts at 0.4 percent, and roughly half the unsupported claims. A model that discloses a problem instead of asserting a result is a materially better tool for long work, and the Quality score of 8 rests partly on it. The second is the legal and finance result: first of 94 on Vals Legal Research Bench and first on its Finance Index are measured on the work itself, and they are the strongest reason a professional in those fields should test Carbon. The review's criticism is about time, verbosity, the price change, the report-quality finding in the six-task build and the skill position, not about the model's honesty.
What this review does not do. It does not score the /swarm preview as a product, because the vendor labels it a preview and the only measurements are the vendor's own. It does not score Carbon for cyber as a separate product, because that variant sits behind the Fairwind allowlist and its evaluations were run without out-of-model guardrails. It does not treat the introductory price as a permanent one or the paid-only boundary as a safety guarantee. And it does not convert a benchmark row into a promise: a model that scores 65.2 on agentic terminal work will fail at agentic terminal work, and the reader who plans for that is the reader the framework is written for.
Review conducted by URC under the CI-First Evaluation Framework, version 1.4.1. Scoring date 2026-11-13. Tool version reviewed: Gemini 4 Carbon (`gemini-4-carbon`, announced 2026-10-28; generally available 2026-11-10). Framework version applied: 1.4.1.
Gemini 4 Carbon: Google's strongest model is not its best buy
Status and Re-check
Status: Active | Last tested: 2026-11-13 (Gemini 4 Carbon, `gemini-4-carbon`, announced 2026-10-28; generally available 2026-11-10) | Re-check: trigger-based (max 6 months)
Active: the tool is current and recommended.
For detailed explanations of the CI-First evaluation terms used in this review, including the Humics Protection Badge and the AI Imposture Risk levels, see the Glossary at the end of this post.
Re-check triggers:
The price change on 2026-12-01. Every rate in this review doubles on that date: input from $2.00 to $4.00, output from $10.00 to $20.00, cached input from $0.10 to $0.20, and Priority from $3.60 and $18.00 to $7.20 and $36.00. The cost per task that carries the Quantity score is built on the introductory rate, and the arithmetic has to be re-run at the standard one.
A new frontier release or a new effort setting on Claude Fable 5.2, Claude Opus 5.5 or GPT-6 Astra. Carbon led the independent index for twelve days and now sits fourth behind three settings of Fable 5.2. The comparison is the most unstable part of this review.
Any change to the three thinking levels, the 65,536-token synchronous output ceiling, or long decode. The Time and Quantity scores rest on how much work one call can carry at a chosen effort level, and the default is medium rather than the setting most independent runs used.
Any change to the Fairwind Program or to the cyber safeguards. The cyber scores were produced without out-of-model guardrails and through a restricted access path, and the threshold position is carried forward from Gemini 4 Argon.
A published incident in a long-horizon agent deployment, including prompt injection, tool-use failure or a false completion report. The vendor's honesty and safety claims are engineering claims, and a claim of this kind is tested by incidents rather than by release notes.
Any change to the paid-only data terms, or the introduction of a free tier or a Contributor tier for Carbon. Carbon has no free route at all, as Argon has none, and that is a data fact before it is a price fact.
A published hallucination or honesty result that contradicts the 11 percent and 92 percent figures. The Quality score and the Skill Illusion rating rest on exactly those two numbers.
A controlled head-to-head of Carbon, Fable 5.2 and Opus 5.5 on cost per completed task at matched effort. No such run exists in public, and the published cost figures are measured on different test setups at different effort settings.
Comments