top of page
Abstract Shapes

INSIDE

PUBLICATIONS

Doubao Seed 2.1 Pro: ByteDance's million-token model, priced in renminbi and measured by nobody yet

7 hours ago
55 min read
Doubao Seed 2.1 Pro cover art: a typed lockup naming BYTEDANCE and Doubao Seed 2.1 Pro over an archive of documents converging into one shaft of light

Status: Active | Last tested: 2026-11-19 (Doubao Seed 2.1 Pro, doubao-seed-2-1-pro-260915, vendor snapshot token 260915) | 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.


Doubao Seed 2.1 Pro Review
Back to the TOC

In this Tool Review


Back to the TOC

Section icon: The name on the price sheet, the two ids a developer can call, and the number that is not a release date.

Doubao Seed 2.1 Pro

The name on the price sheet, the two ids a developer can call, and the number that is not a release date


Doubao Seed 2.1 Pro is a hosted model, so the identity that matters is the model id a request carries. The vendor uses three different strings for one product family, and the difference between them decides whether a call succeeds and how much context the call gets.


Name

What it is

Where it appears

Doubao Seed 2.1 Pro

The product name this review scores. One member of ByteDance's Doubao Seed family, served on Volcano Engine (火山方舟, Ark)

The model list, the model price sheet and the vendor's own documentation navigation

`doubao-seed-2-1-pro-260915`

The current callable Model ID. A 1,024K context window, a 1,024K maximum input, a 256K maximum answer with a 4K default, and a 256K maximum thinking chain

The model list, and every official code sample on the text-generation, image-understanding, deep-thinking and product-overview pages read on 2026-11-19

`doubao-seed-2-1-pro-260628`

An earlier callable snapshot, still listed under the model list's "other models" section. A 256K context window and a 256K maximum input

The model list, under "other models"

`doubao-seed-2.1-pro`

The vendor's billing and family name on the price sheet. It is not a Model ID in the model list, and no official code sample passes it

The model price sheet, one row, one price band of [0, 1024]K input tokens

Doubao Seed 2.1 Turbo, Doubao Seed 2.1 Lite, Doubao Seed 2.0 Pro, Doubao Seed Evolving

Four different models in the same catalogue, each with its own row and its own price. They are not variants of this product and this review does not score them

The model list and the price sheet


This is a new review, not a re-score and not a family batch. No Doubao or Seed record existed in the Tools collection, no Doubao post existed among the published Tools posts, and no local review document carried the name before this one was written. The four sibling models above appear in this document only as the alternatives a buyer compares against.


The absence of a published release date is a fact about the vendor's pages, not an omission by this review. The model list, the price sheet, the product overview and the deep-thinking guide were read on 2026-11-19 and none of them states a calendar release date for this model. The 260915 and 260628 suffixes are the vendor's own snapshot tokens, and this review does not convert a snapshot token into a date claim. A reader who needs a date should take it from the vendor's account team rather than from an inference made here.


What this review is not. It is not a review of the Doubao consumer app, of the Doubao Assistant, of Seedance video generation or of Seedream image generation. It is not a review of the Volcano Engine platform as a whole, although account, region and payment questions appear where they decide whether a reader can use the model at all. It does not score Doubao Seed 2.1 Turbo, Lite or 2.0 Pro, and it carries none of their numbers into this one.


Back to the TOC


Section icon: Tool Snapshot.

Doubao Seed 2.1 Pro

Tool Snapshot


Doubao Seed 2.1 Pro, ByteDance's long-context flagship on Volcano Engine: a 1,024K window, a 256K answer ceiling and thinking on by default, priced in renminbi with a single band across the whole window and no independent measurement published anywhere this review could find.


Category: Applied AI / Large Language Model, built for long-document analysis, multilingual text work, agentic tool use and GUI task handling. It is a hosted, closed-weight model, reachable through ByteDance's Volcano Engine platform.


Primary use cases:


  • Whole-corpus reading in one call. A 1,024K-token input window and a 256K maximum answer mean a contract set, a codebase slice or a reporting pack can be submitted without chunking, and the answer can be long enough to be a deliverable rather than a summary.

  • Agentic tool use and structured output. The model list records tool calling, structured output and GUI task handling as supported capabilities, and the platform documents a GUI Agent surface. This is the configuration in which the model is asked to operate a workflow rather than answer a question.

  • Multimodal document work. The model list records multimodal understanding, and the image-understanding guide uses doubao-seed-2-1-pro-260915 in its own samples. Images and documents can enter the same window as text.

  • Reasoning-heavy analysis with a visible trace. Deep thinking is supported and enabled by default, with a 256K maximum thinking chain, and the platform documents how to disable it per request.

  • Batch and low-priority processing. The price sheet publishes a batch rate and a low-priority rate at half the regular input and output price, which is the route for work that can wait.


What it is not. It is not a model with any independent measurement behind it: no third-party index row and no published benchmark table appears in the vendor's documentation. It is not open weights, not self-hostable and not inspectable from outside the vendor's service, and no parameter count is published. It is not a model sold on a global rate card: the documentation is published in Chinese and the prices are in renminbi. It is not a teaching product, and nothing in it is built to move a capability from the model to the person. It is not an image, video or audio generator, and audio input is not priced for this model at all.


Pricing, as published on 2026-11-19 (vendor currency, renminbi):


Line

Doubao Seed 2.1 Pro

Input, non-audio, regular online inference

CNY 6.00 per million tokens

Cached input, non-audio

CNY 1.20 per million tokens

Cache storage

CNY 0.017 per million tokens per hour

Output, thinking tokens included

CNY 30.00 per million tokens

Input band

One band: input length [0, 1024]K tokens. No long-context multiplier is published

Batch inference

CNY 3.00 input, CNY 1.20 cached, CNY 15.00 output per million tokens

Low-priority online inference

CNY 3.00 input, CNY 1.20 cached, CNY 15.00 output per million tokens

Low-latency online inference

Not offered for this model: the low-latency table lists the Lite, Turbo and 2.0 series and does not carry a 2.1 Pro row

Audio input

Not priced: the audio input column carries no rate for this model

Currency note

The vendor quotes renminbi only. This review publishes no dollar figure, because a converted figure without a named rate and date is a number a reader cannot check


The price argument needs one sentence of arithmetic, because the tiers are not small print. The batch and low-priority rates are exactly half the regular input and output rates, and the cached-input rate is one fifth of the regular input rate. A team that sends long, stable preambles and processes overnight work on the batch route pays a materially different bill from a team that sends the same tokens interactively, and both are published on the same page. Thinking tokens are billed at the output rate, so the default reasoning setting is a cost decision before it is a quality decision.


Official links:



At a Glance Dashboard


Field

Value

Category

Applied AI / Large Language Model (long-context document work, agentic tool use, multimodal input, China-hosted)

CI-First Benefit Score

5.0 / 10 (CI-First Positive)

Sub-scores

Time 6 / Quantity 7 / Quality 4 / Skill 3

CI-First Profile

Primary: Co-Worker and Assistant (level 2). Secondary: Co-Researcher, Analyst and Tester (level 4), Professor, Coach and Tutor (level 3)

Humics Protection

Humics-Neutral (-1 / +3)

AI Imposture Risk

Medium overall, with all three traps at Medium

Status

Active

Last tested

2026-11-19

Pricing

CNY 6.00 per million input tokens, CNY 1.20 cached input, CNY 30.00 per million output tokens, one band across the 1,024K window; batch and low-priority at half rate

Platforms

Volcano Engine (火山方舟, Ark): Chat API, Responses API and Messages API, plus the platform console

User sentiment

No independent measurement and no review-platform listing were found on 2026-11-19; see What Users Say for what that does and does not mean


Back to the TOC


Section icon: The Problem.

Doubao Seed 2.1 Pro

The Problem


The work that needs a very large context window is rarely one question. It is a set of documents that only makes sense together: forty supplier contracts with annexes, a codebase slice with its tests and its build files, a year of board packs, a regulatory filing set in two languages. A model with a 128K window forces that work into chunks, and chunking is where the analysis breaks, because the sentence that decides the answer is in the annex of a document the chunking step put in another call.


The second problem is price shape. A long-context request is mostly input, and an input rate multiplied by a very large token count is the bill. Models that charge a multiplier above a certain input length make a reader choose between a complete document set and a budget. A single band across a million tokens changes that decision, and a cached-input rate that is a fifth of the input rate changes it again for the preambles a team sends every day.


The third problem is evidence. A buyer outside mainland China has, at the time of writing, no independent measurement of this model to reason from: no third-party index row, no published head-to-head on a fixed task set, and no review-platform record. That does not make the model bad. It makes the purchase a decision made on vendor documentation, a trial of your own, and the terms of a platform account, and a reader deserves to be told which of those three the answer rests on.


Back to the TOC


Section icon: The Outcome.

Doubao Seed 2.1 Pro

The Outcome


What a reader gets from this model. A hosted flagship with a 1,024K input window, a 256K answer ceiling, a 256K thinking chain, multimodal input, tool calling, structured output and a GUI agent surface, priced at CNY 6.00 per million input tokens and CNY 30.00 per million output tokens, with batch and low-priority routes at half and cached input at one fifth. For a document-heavy workflow, that is a real capability: sets that had to be split can be read whole, and a deliverable can be long enough to be used.


What a reader does not get. A measured quality claim. The vendor publishes limits and capabilities, not benchmark results in its documentation, and no independent house in the usual set publishes a row for this model. The score reflects that absence rather than pricing the architecture as if it were a result. The band stays CI-First Positive: 5.0 is a real recommendation for the work the model is built for, with disciplined use and a check the reader owns.


What changes in practice. Three habits carry most of the value. Pin the versioned model id, because the family name on the price sheet is not callable and the snapshot decides whether the window is 256K or 1,024K. Decide the thinking setting before the run, because thinking is on by default and its tokens are billed at the output rate. Choose the route, regular, batch or low priority, because the same tokens cost half on the routes that can wait.


Back to the TOC


Section icon: Who Should Use Doubao Seed 2.1 Pro.

Doubao Seed 2.1 Pro

Who Should Use Doubao Seed 2.1 Pro


Learner type

Difficulty

Typical ROI

Career path

Students (Bachelor, Master)

Intermediate

A long-context model for a literature set, a thesis corpus or a codebase slice, where the gain is that the whole set fits in one call and the risk is that a fluent summary of a paper the student has not read looks like understanding. The 1,024K window is the teaching-relevant feature and the Skill Illusion is the teaching-relevant danger: a student who can produce a comparison table without reading the sources has learned to produce tables. Access requires a Volcano Engine account, which is a real step for a student outside China.

UIT (Technology, AI, Data Science) tracks for context engineering, evaluation design and cost arithmetic. UIB (Business Management, Entrepreneurship) for the commercial reading of a rate card with three routes. No U365 credential is mapped to this review.

Professionals (career upskilling)

Intermediate

The strongest case: contract sets, filing sets, codebase slices and reporting packs read whole, with a 256K answer ceiling that allows a usable deliverable rather than a summary, and a half-price batch route for work that can wait. The limit is evidential rather than technical: there is no independent measurement to calibrate trust against, so the professional's own check is the only check.

UIT (Technology, AI, Data Science) for tool calling, structured output and agent supervision. UIB (Business Management, Entrepreneurship) for procurement and vendor-risk reading of a China-hosted service.

Everyone (lifelong learners)

Beginner to Intermediate

A capable assistant for long documents and multilingual reading, reached through the API rather than a consumer subscription. The gain is a first pass over material that would otherwise go unread; the cost is an account, a credit balance and a habit of checking. It is the wrong tool for a one-line question, because thinking is on by default.

LIPS and CARE 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. The model is an API product on a Chinese cloud platform, and a team deciding on it starts with four questions.


  • Team deployment. One model id, one key, one endpoint shape, and a decision the team must write down: which snapshot is pinned. The current snapshot carries a 1,024K window and the earlier one a 256K window, on the same price row, so an unpinned deployment is a deployment whose capacity can change under it. The platform also exposes a console, a batch route and a low-priority route, and the team's job is to assign each workload to a route rather than leave it to habit.

  • Cost at team scale. Three lines move independently: the input rate, the output rate and the thinking setting. The input rate is CNY 6.00 per million tokens and falls to CNY 3.00 on batch and low priority; cached input is CNY 1.20; output is CNY 30.00 and carries the thinking tokens. A team that sends a large stable preamble on every call without caching pays the full input rate on text it has already sent, and a team that leaves thinking on for extraction work pays the reasoning rate for tasks that did not need reasoning.

  • Data and compliance. The service is hosted on Volcano Engine under the platform's terms, the documentation is published in Chinese, and the prices are in renminbi. A team outside mainland China has to settle account eligibility, contracting entity, data location and payment before it settles anything about the model. A legal, tax or financial analysis produced by the model remains a draft that a qualified person reviews, and the absence of independent measurement makes that review more important rather than less.

  • 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: Intermediate. The model is easy to call and the hard part is scope: deciding what belongs in a million-token window, writing the instruction that keeps the answer inside the job, choosing whether thinking is worth its tokens, and knowing which part of the output you can check yourself. None of that is needed to get an answer, and all of it is needed to rely on one.


Prerequisites. A Volcano Engine (火山方舟, Ark) account with the model activated, an API key, and a decision about the snapshot id the code pins. Working knowledge of prompt structure and of verification practice, because the output reads as finished whether or not it is correct. A Chinese-reading colleague or a translation step if the team does not work in Chinese, because the documentation set is published in Chinese.


Typical time to first result: Under thirty minutes for a first successful call, most of it account setup rather than code. The call itself needs three decisions: the snapshot id, the thinking setting and the maximum answer length.


Typical time to competence: Ten to twenty hours of active use. Two habits carry most of the value: assigning each workload to a route before sending it, and pairing every long-context answer with a check you already own.


Who should wait. Anyone whose work is short requests, because thinking is on by default and the setup cost never repays itself. Anyone whose procurement cannot contract with a Chinese cloud platform, because the model is not available on another one. Anyone who needs an independently measured quality claim before adopting, because none is published. And anyone who needs image, video or audio generation, because this model produces text.


Back to the TOC


Section icon: U365 Institutes Alignment.

Doubao Seed 2.1 Pro

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 the product's distinctive surface is context engineering at a million tokens with tool calling and structured output behind it. 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. Context engineering: a 1,024K window is not a prompt, it is a selection problem, and deciding what to include, what to cite and what to leave out is the discipline the window makes visible. Structured output and tool calling: the model list records both, and using them well means writing a schema, bounding the tool set and checking the returned structure rather than reading it. Cost engineering: three published routes with a 2:1 ratio and a cached-input rate at one fifth of input is arithmetic a cohort can practise on live numbers. Evaluation design: a model with no independent measurement is the clearest case a cohort will meet of the difference between a vendor claim and a measured result, and building a small fixed task set of its own is the exercise.

  • 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 documentation set is in Chinese, which adds a reading burden for a cohort working in English. The catalogue's nearest anchor is context-engineering instruction, carried by three published programmes (NOVEC - Acculturation IA - Jour 3/7 - Context Engineering, Superhuman Expert with AI - Diploma and Superhuman@Work - Certificate), and no published programme's title or description names structured output, tool calling, cost engineering or evaluation design. Adjacent anchors, not assessment homes. No U365 credential is mapped here.


UIB (Business Management, Entrepreneurship)


  • Rating. Medium

  • Why. Two decisions the product forces that are commercial before they are technical. Procurement of a China-hosted service: account eligibility, contracting entity, data location, payment currency and the terms that govern all four are business decisions, and reading them is a finance and vendor-risk skill. And cost modelling across routes: CNY 6.00 against CNY 3.00 on input, CNY 1.20 on cached input, CNY 30.00 on output with thinking included, and a single band across the window, which together decide whether a long-document workflow is affordable at volume. A business case built on the regular rate alone is a case built on the most expensive route.

  • The limit that holds the row. The model supplies no management method and measures no return of its own. It publishes no completion rate and no time-per-task figure, so a business case has to be built and tested by the unit. No published programme's title or description names procurement or vendor risk. No U365 credential is mapped here.


UIC (Digital Communication, Marketing)


  • Rating. Low to Medium

  • Why. One competency that matters and is easy to miss: reading a vendor claim against the document that supports it. This model is a worked example, because a 1,024K window, a single price band and a half-price batch route are all true and none of them is a quality claim, and a communication practitioner who can separate the two is doing the job the profession claims. The writing surface is real as well: a long-context model can hold a brand guide, a set of past campaigns and a brief in one window, which is the shape of a drafting task rather than a chat.

  • 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 by default, 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 design contact is input-side and interface-side rather than visual authorship. The model reads images and documents alongside text, and the platform documents a GUI Agent surface, so a designer can use it to read a screen set, compare states or turn a described interaction into structured output. Studying how a model reads a layout is useful preparation for judging what a design system has to make explicit.

  • The limit that holds the row. The model generates no images and evaluates no design against a brief. It produces text and structured data, so the visual surface is input plus specification rather than visual authoring. No layout, interaction or motion competency is assessed by it. No credential is mapped here.


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 large window is: scope is a planning decision before it is a prompt decision, and the UP-Context-Engineering Method's eight layers are what keep a million-token request inside the job it was given. USER-PERSONA and FULL-CONTEXT decide what goes into the window; AI-EXACT-TASK and EXPECTED-FORMAT-AND-CONSTRAINTS decide what comes out; the verification habit is LIPS and CARE in practice. The batch and low-priority routes are a planning discipline as well, because work that can wait should be planned to wait. That choice is also a recurring cost decision in the Career and Finance domain of ULM, where the EVA engine (Explore, Visualize, Action Plan) turns it into a plan, and cost per completed task is CI-First cost engineering and the independent evaluator's metric, not a ULM method. UNOP has no direct fit: the model is not built to teach, and its gains come from capacity and disciplined execution rather than from a 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 record what was asked, which snapshot answered, at what thinking setting and on which documents has no account of the run, and nothing in the product writes that account for them.


Back to the TOC


Section icon: How Doubao Seed 2.1 Pro Works.

Doubao Seed 2.1 Pro

How Doubao Seed 2.1 Pro Works


Inputs. Text, and images or documents through the multimodal understanding surface, submitted through the platform's Chat API, Responses API or Messages API. The published limits are 1,024K tokens of context, a 1,024K maximum input and a 256K maximum answer, with a 4K default answer length and a separate 256K maximum thinking chain. A request that carries more than the answer ceiling needs splitting or a longer-output pattern, and the default of 4K is the setting most likely to truncate a first attempt.


Reasoning. Deep thinking is a supported capability and is enabled by default on a call. The platform documents two ways to turn it off: thinking.type=disabled on the Chat API and reasoning.effort=minimal on the Responses API. Thinking tokens are part of the output bill, and the thinking chain has its own 256K ceiling. A reader who sends an extraction task without setting this is paying a reasoning rate for work that may not need reasoning.


Capabilities beyond text generation. The model list records four that change what a workflow can look like: multimodal understanding, tool calling, structured output and GUI task handling. Structured output is the one to use when a downstream system reads the answer, because a schema turns a prose response into a record. GUI handling is the one to read the platform guide for before planning, because it changes the failure modes rather than the price.


What it does not expose. No parameter count, no architecture description, no weights, no self-hosting route and no published training-data statement appear in the vendor's documentation. The model holds no state between calls: memory, retrieval and the record of what was asked belong to the application that wraps it, not to the model.


Rate limits. The model list publishes 500 maximum requests per minute and 1,000,000 maximum tokens per minute, and labels them non-rigid, affected by platform load and by call pattern. A team planning a large batch should read those two numbers before promising a throughput figure to anyone.


Where it runs. Volcano Engine (火山方舟, Ark). The documentation is published in Chinese, the price sheet is in renminbi, and the model ids are Volcano Engine model ids. That is the deployment context a reader should carry into any planning conversation, because it decides contracting, data location and payment before it decides anything about prompts.


Back to the TOC


Section icon: Getting Started with Doubao Seed 2.1 Pro.

Doubao Seed 2.1 Pro

Getting Started with Doubao Seed 2.1 Pro


A fifteen-minute checklist, in the order the platform requires.


  • Create the account and activate the model. A Volcano Engine account with Ark (火山方舟) activated and this model enabled. The console is where the model is enabled and where the price sheet's routes are selected.

  • Create an API key and store it outside the code. The key belongs in a secret store, not in a notebook and not in a repository.

  • Pin the snapshot id in the request. Use doubao-seed-2-1-pro-260915 for the 1,024K window. Do not send doubao-seed-2.1-pro: that is the billing name on the price sheet, not a Model ID in the model list. Pinning the snapshot is what makes the capacity of the deployment a decision rather than a coincidence.

  • Set the answer ceiling before the first long call. The default answer length is 4K. A 256K ceiling is available and the request has to ask for what it needs.

  • Decide the thinking setting per task. Leave it on for analysis and planning. Turn it off for extraction, classification and reformatting, using thinking.type=disabled or reasoning.effort=minimal depending on the API surface. Write the decision down, because the default is on.

  • Choose the route. Regular online inference for interactive work, low priority for work that can wait, batch for a job with no deadline. The input and output rates on the waiting routes are half the regular rates.

  • Send one small call that proves the wiring. A short prompt with a known answer, checking the returned model id and the token usage fields, before any long document is sent.

  • Send one real document and measure it. Take the usage numbers from the response rather than estimating them, and multiply by the rate for the route actually used. That number, not the list price, is the unit cost of the workflow.

  • Write the log line. Snapshot id, thinking setting, route, token counts and the document set. The model keeps no record, so the record is yours.


What to do when the first long call fails. Three causes cover most of it. A model id that is the family name returns an error rather than an answer, so check the id first. An answer that stops early is usually the 4K default ceiling rather than the window. A bill that looks wrong is usually thinking left on, or a route other than the one assumed.


Back to the TOC


Section icon: Real Workflows.

Doubao Seed 2.1 Pro

Real Workflows


Three workflows, each chosen because the model's published capacity and a check the reader already owns sit in the same place. Each sample prompt uses the UP-Context-Engineering Method's eight layers in method order, with the constraints and the data limits inside EXPECTED-FORMAT-AND-CONSTRAINTS.


Workflow 1: Reduce a contract set to a clause comparison you can defend


When to use it: you have a set of agreements that only makes sense together, the window can hold them, and a qualified person will read the result. This is the workflow the 1,024K window was built for and the one where a fluent answer is most dangerous, because a comparison table with a plausible cell in every row looks finished.


This sample prompt uses the UP-Context-Engineering Method and the CI-First doctrine. Read the method at https://www.university-365.com/post/the-up-method-university-365-prompting-the-new-gold-standard-for-prompt-engineering, the doctrine at https://www.university-365.com/post/discover-the-ci-first-approach, learn both in a U365 online program at https://www.university-365.com/programs, and request a U.Coach for orientation at https://www.university-365.com/ucoach.


USER-PERSONA: My UP-Context-Engineering USER-PERSONA file is the file for this layer, attached. It is my psychographic file concatenated with the six ULM life-domain files. The parts that matter here: I am the reader who signs the recommendation, I am not a lawyer, and I do not accept a summary I cannot trace to a clause. If the file is not attached, ask for it before you start.
FULL-CONTEXT: The FULL-CONTEXT set for this task, added and not merged: BUSINESS-PERSONA for [the organisation], PROJECT-DESCRIPTION for [the deal or the review], the agreement set [file names], the clause list I care about [list], the governing law of each agreement [list], and the questions the business has already asked [list]. The documents are the only authority and my clause list is the only scope. If a document is missing or unreadable, name it and stop.
AI-PROFILE-ROLE-EXPERTISE: Act as a Co-Researcher, Analyst & Tester, in the role of a contract analyst who compares operative clauses across agreements, with expertise in clause structure, defined terms and the difference between a recital and an obligation.
USER-FINAL-GOAL: A comparison table my lawyer can check clause by clause, with the citation for every cell, so the review meeting decides rather than discovers.
AI-EXACT-TASK: Build one table, one row per agreement and one column per clause in my list.
AUDIENCE: Me, and a qualified lawyer who will check every cell against the documents. The lawyer has not seen your reasoning and will not accept a summary as evidence.
EXPECTED-FORMAT-AND-CONSTRAINTS: Format: (1) the table, (2) one paragraph per agreement naming what the table does not capture, (3) the list of documents you could not read in full. Constraints: quote the operative words and cite the document and the clause number in every cell; never merge two documents into one row; write Silent, not a probable answer, in a cell where an agreement is silent; mark any cell you are unsure of as UNVERIFIED with the reason; do not give legal advice or a recommendation; if the set exceeds the input window, stop and tell me which documents to drop rather than truncating silently; thinking is on by default and I am paying the output rate for it, so keep the visible answer to the table and the two paragraphs. Data limits: the agreement set is cleared for this model at the level I state in the FULL-CONTEXT set; no privileged material, personal data or unreleased commercial term goes into a surface I have not cleared.
EXAMPLES: The clause list format from the last review is attached as the column example, and one agreement with its citations marked is attached as the cell example.

Before you use the output, run the check yourself: open three cells at random and find the quoted words in the document, confirm that every silent cell says Silent, and read the paragraph on what the table does not capture. The decision is yours, per CI-First.


What the model does well here: capacity. A set that had to be split into six calls fits in one window, so the cross-document question is asked once and the answer is written against the whole set. The published capability that matters is the input ceiling, not a quality claim, and this review scores it as capacity.


VERIFICATION CHECKLIST for Reduce a contract set to a clause comparison you can defend:


  • ☐ Multi-Model Check: run the same clause list through a second model (Claude Fable 5.2 or Gemini 4 Neon) and compare the cells that disagree. A disagreement is a cell to read yourself, not a cell to average.

  • ☐ External Source: open every quoted clause in the source document. A quote that cannot be found is a fabricated cell, whatever the rest of the table looks like.

  • ☐ Human Review: a qualified lawyer reads the table and the silence paragraph before the review meeting.

  • ☐ CI-First Test: can you defend each row without the transcript, from the documents? If not, the table is not evidence yet.


Workflow 2: Triage a codebase slice with a tool, then read the diff


When to use it: a failing area spans more files than you want to open, the slice fits the window, and you can run the tests yourself. The end state is a diff you can read and a command you can re-run.


This sample prompt uses the UP-Context-Engineering Method and the CI-First doctrine. Read the method at https://www.university-365.com/post/the-up-method-university-365-prompting-the-new-gold-standard-for-prompt-engineering, the doctrine at https://www.university-365.com/post/discover-the-ci-first-approach, learn both in a U365 online program at https://www.university-365.com/programs, and request a U.Coach for orientation at https://www.university-365.com/ucoach.


USER-PERSONA: My UP-Context-Engineering USER-PERSONA file is the file for this layer, attached. It is my psychographic file concatenated with the six ULM life-domain files. The parts that matter here: I own this merge and I defend the diff line by line. If the file is not attached, ask for it before you start.
FULL-CONTEXT: The FULL-CONTEXT set for this task, added and not merged: BUSINESS-PERSONA for [the organisation], PROJECT-DESCRIPTION for [repository], the repository at [path] on branch [branch], the failing command [command], the failing tests [names], the call path I suspect [files], and the test file whose assertions must not change. The repository is the only code authority and the test run is the only proof. If the repository path or the failing command is missing, name it and stop.
AI-PROFILE-ROLE-EXPERTISE: Act as a Co-Researcher, Analyst & Tester, in the role of an engineer tracing a failure across module boundaries, with expertise in call-path reading, minimal diffs and test assertions.
USER-FINAL-GOAL: The suite passes on a clean checkout, the diff is small enough to read in one sitting, and I can explain the cause without the transcript.
AI-EXACT-TASK: Trace the failure to its cause, make the smallest change that makes the failing tests pass, and explain the call path in five sentences.
AUDIENCE: Me, and a reviewer who reads the diff without running the session.
EXPECTED-FORMAT-AND-CONSTRAINTS: Format: (1) the call path in five sentences, (2) a unified diff, (3) the exact command to re-run, (4) what you did not verify. Constraints: stay inside the file list; use the repository's own test command rather than inventing one; do not reformat or refactor outside the failing path; do not add a dependency; do not weaken, skip or delete a test; do not claim a run you did not perform; if the cause is outside the files in scope, stop and name the file you need; the answer ceiling is 256K and thinking is on by default, so if the diff and the reasoning will not fit, say so before you start; do not state a knowledge cutoff you were not given, and treat the repository as the only source for anything after my last supplied document. Data limits: the repository is cleared for this model at the level I state in the FULL-CONTEXT set; no credential, private key or customer record leaves my control.
EXAMPLES: The last fix in this area that passed the suite is attached as the diff example, and the existing test file is the assertion-style example.

Before you use the output, run the check yourself: re-run the exact command in a clean checkout, read the diff line by line, confirm that no test was skipped or weakened, and explain the cause without the transcript. The decision is yours, per CI-First.


What the model does well here: tool calling and structured output in one window. A slice of a repository plus its test names plus a declared tool set is a task shape the platform documents, and the 256K answer ceiling is enough for a diff plus an explanation. The published limit is not a quality claim, and the review does not treat it as one.


VERIFICATION CHECKLIST for Triage a codebase slice with a tool, then read the diff:


  • ☐ Multi-Model Check: run the diagnosis through a second model and compare the cause, not the diff.

  • ☐ External Source: run the test command yourself in a clean checkout and confirm the failing tests pass and no assertion changed.

  • ☐ 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 call path and defend the fix without the transcript?


Workflow 3: Batch-extract a document archive into records


When to use it: the work is volume, the schema is stable, and nobody needs the answer today. This is the workflow where the half-price batch route matters and where the thinking setting is the difference between a sensible bill and a strange one.


This sample prompt uses the UP-Context-Engineering Method and the CI-First doctrine. Read the method at https://www.university-365.com/post/the-up-method-university-365-prompting-the-new-gold-standard-for-prompt-engineering, the doctrine at https://www.university-365.com/post/discover-the-ci-first-approach, learn both in a U365 online program at https://www.university-365.com/programs, and request a U.Coach for orientation at https://www.university-365.com/ucoach.


USER-PERSONA: My UP-Context-Engineering USER-PERSONA file is the file for this layer, attached. It is my psychographic file concatenated with the six ULM life-domain files. The parts that matter here: I run the archive and I answer for every record I publish. If the file is not attached, ask for it before you start.
FULL-CONTEXT: The FULL-CONTEXT set for this task, added and not merged: BUSINESS-PERSONA for [the organisation], PROJECT-DESCRIPTION for [the archive], the document set [path or manifest], the schema [field list with types and allowed values], the date range [range], and the two documents already extracted by hand as the reference records. The schema is the only contract and the hand-extracted pair is the only style authority.
AI-PROFILE-ROLE-EXPERTISE: Act as a Co-Worker and Assistant, in the role of an extraction operator producing records against a fixed schema, with expertise in field mapping, null handling and evidence marking.
USER-FINAL-GOAL: A record set the downstream system can load without cleanup, with every field either extracted from the document or explicitly null, so the archive stops being a reading queue.
AI-EXACT-TASK: For each document, return one record matching the schema.
AUDIENCE: A downstream system that will reject a malformed record, and me, who will spot-check the output before it loads.
EXPECTED-FORMAT-AND-CONSTRAINTS: Format: JSON only, one object per document, keys exactly as the schema defines, no prose before or after. Constraints: extract only what the document states; return null in a field the document does not state rather than a probable value; mark the page or section for every non-null field; no merged documents, no reformatting of dates beyond the schema's stated form; if a document is unreadable, return a record with the error field rather than skipping it silently; thinking off for this task; batch route, because nothing here is interactive; the input window is 1,024K and the answer ceiling is 256K, so split a document that will not fit rather than truncating it. Data limits: the archive is cleared for this model at the level I state in the FULL-CONTEXT set; no personal data category outside the schema, and no document from a collection I have not cleared.
EXAMPLES: The two hand-extracted reference records are attached, one clean and one with nulls, as the format and null-handling examples.

Before you use the output, run the check yourself: validate every record against the schema, open ten at random against their documents, and count the nulls per field. A field that is never null is a field the model is guessing. The decision is yours, per CI-First.


What the model does well here: structured output at volume on a half-price route. A stable schema, a large window and a batch price are the shape of an extraction job, and the model list records structured output as supported. The cost argument is the published route ratio, not a throughput claim.


VERIFICATION CHECKLIST for Batch-extract a document archive into records:


  • ☐ Multi-Model Check: run twenty documents through a second model against the same schema and compare the records that disagree.

  • ☐ External Source: validate every record against the schema and open ten at random against their source documents.

  • ☐ Human Review: a person signs the load, and reads the per-field null counts before it ships.

  • ☐ CI-First Test: can you explain, for any record, where each field came from?


Back to the TOC


Section icon: Strengths, Limits, and AI Imposture Risk.

Doubao Seed 2.1 Pro

Strengths, Limits, and AI Imposture Risk


Strengths


Time


  • Strength. Whole-set reading in one call, which removes the chunk-and-reconcile loop that dominates long-document work.

  • Evidence. A 1,024K-token input window and a 256K maximum answer on the current snapshot, published on the model list. The saving is structural rather than measured: it comes from not splitting the set, not from a published speed figure. Held at 6 because thinking is on by default, no latency measurement is published, and account setup plus Chinese-language documentation is real overhead for a team outside China.


Quantity


  • Strength. A large output ceiling and three price routes, which together raise how much usable work one budget produces.

  • Evidence. A 256K maximum answer with a 4K default, a separate 256K thinking chain, batch and low-priority routes at exactly half the regular input and output rates, and cached input at CNY 1.20 against CNY 6.00. Held at 7 because no throughput measurement is published and the published rate limit of 1,000,000 tokens per minute is labelled non-rigid.


Quality


  • Strength. Unmeasured. The capability set is broad; the quality claim is the vendor's and no independent measurement exists.

  • Evidence. The model list records deep thinking, multimodal understanding, tool calling, structured output and GUI task handling. No third-party index row and no published benchmark table were found on 2026-11-19. Scored at 4, the framework's measured-absence position: an architecture is not a result.


Skill


  • Strength. Marginal, and honestly marginal.

  • Evidence. No teaching mode is claimed. The visible thinking chain is a reasoning surface a reader can study, which is a small real benefit, and the documentation is in Chinese, which narrows the learning surface for a reader working in English. Clause 5.2.3-a returns a null with a boundary rather than a mechanism that transfers capability. Scored conservatively.


Limits


  • No independent measurement exists. No third-party index row, no published head-to-head on a fixed task set and no review-platform record were found on 2026-11-19. A reader cannot calibrate the model against a peer before buying, and the Quality score reflects that absence rather than a judgment about the model's capability.

  • The family name is not callable. doubao-seed-2.1-pro is the billing name on the price sheet. The callable Model IDs are doubao-seed-2-1-pro-260915 and doubao-seed-2-1-pro-260628, and the two do not have the same window.

  • The snapshot decides the capacity. 1,024K on 260915 against 256K on 260628, on the same price row. An unpinned deployment is a deployment whose window can change underneath it, and a pinned one is a deployment that will eventually be superseded.

  • Thinking is on by default and billed as output. A request that does not set the thinking parameter pays the output rate for reasoning it did not ask for. The two documented switches differ by API surface, so the switch is a code decision rather than a console setting.

  • The answer default is 4K. The 256K ceiling is available and has to be requested. A first long call that stops early is usually the default, not the window.

  • The documentation and the rate card are Chinese and renminbi only. The vendor publishes no dollar price and no English equivalent. A buyer outside mainland China has to settle account eligibility, contracting entity, data location and payment before the model question.

  • No low-latency tier for this model. The low-latency table carries the Lite, Turbo and 2.0 series and does not carry a 2.1 Pro row, so a latency-sensitive workload has no premium route here.

  • No audio input price. The audio input column is empty for this model, so a speech workflow is not a configuration of this product.

  • Nothing is inspectable from outside. No weights, no parameter count, no architecture description and no self-hosting route.

  • No release date is published. The two snapshot tokens are the vendor's own version markers, and this review does not turn them into a calendar claim.


AI Imposture Risk


Time Illusion


  • Rating. Medium

  • Evidence. The saving is real when the alternative is chunking: one call over a set replaces several calls plus a reconciliation step. The overhead is real as well: thinking on by default, a 4K answer default that has to be raised, account setup and Chinese-language documentation, and no published latency figure to plan against. A reader who assumes a million-token window means a fast answer has confused capacity with speed.


Quantity Illusion


  • Rating. Medium

  • Evidence. The volume is the product: a 256K answer ceiling, three routes, and structured output that turns a long response into records. The failure mode is that a table with a plausible cell in every row, or a record set with no nulls, reads as complete. Nothing in the published material lets a reader calibrate how often that completeness is real, which is exactly why the checks in Real Workflows are written as counts and random samples.


Skill Illusion


  • Rating. Medium

  • Evidence. Expert-looking contract analysis, code diagnosis and extraction produced for readers who may not be able to evaluate it, at a window size that invites dumping the whole set in. The mitigation is structural: the model is a hosted API with no teaching surface and no memory, so a reader who runs the verification habits keeps the judgment. The risk rises wherever the output is accepted because it is long and well formatted, and the absence of independent measurement removes the external signal a reader would otherwise use.


Overall Imposture Risk: Medium. All three traps are Medium, which the framework's arithmetic places at Medium: one, two or three traps at Medium. The band-precedence sentence does not change the outcome, because the Medium band names this exact trap set and no second rating competes with it. The rating would move up if a reader routinely shipped long-context output without a check; it would move down only with independent measurement showing the output holds up.


Framework v1.2 clause note


Clause 5.2.3-a, agent-authored procedural memory: NULL, recorded with its boundary. The model is a hosted API that holds no state between calls and writes nothing on the user's behalf. The platform's GUI Agent and tool-calling surfaces execute within a session the application defines; they do not create or revise the user's skills, memory stores or standing instructions. No documented mechanism ships by default by which this model writes procedural memory for a reader, so the clause returns a null rather than applying. The live condition is specific: any wrapper that lets the model author a skill file, a memory store or a standing instruction on the reader's behalf re-opens the clause at the Medium floor, and any wrapper that lets it revise that memory during use without a per-write decision reaches High.


Clause 4.2-a, agent-mediated conversation: NULL for the model as a generator. Neither erosion condition is met. The model does not speak in a person's name in a human-facing channel by default, and nothing in it 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 Skill Illusion still rates Medium is not a clause mechanism; it is the plain fact that a million-token window invites a reader to hand over a document set whose correctness they cannot check, and that no independent measurement exists to tell them how often the answer holds.


Back to the TOC


Section icon: Section 7c: The snapshot id is the product, and nothing independent has measured it.

Doubao Seed 2.1 Pro

Section 7c: The snapshot id is the product, and nothing independent has measured it


A Section 7c finding is recorded where a supplier's governance, terms or commercial shape has a real result for the reader. The items below were checked against primary sources, and they produce findings a buyer should carry. They are stated as facts of record, and the difference between a vendor statement and a measurement is kept intact.


The same price row covers two products with different capacity, and the model id is the switch. The price sheet lists one row, doubao-seed-2.1-pro, at CNY 6.00 input, CNY 1.20 cached and CNY 30.00 output per million tokens, under one input band of [0, 1024]K. The model list publishes two callable ids under that family name: doubao-seed-2-1-pro-260628 with a 256K context window and a 256K maximum input, and doubao-seed-2-1-pro-260915 with a 1,024K context window and a 1,024K maximum input. A team that pins the earlier snapshot and budgets from the price sheet has bought a quarter of the capacity it planned for, and nothing on the price sheet says so. The finding is not that the vendor is concealing anything: the model list states both rows plainly. The finding is that the identity a request carries is the only place the difference is expressed, so a procurement or architecture document that records the family name has recorded the wrong fact.


No independent measurement exists, and that is the largest single fact in this review. The vendor publishes limits, capabilities and prices on the pages read, and does not publish a benchmark table there. The usual independent houses were checked on 2026-11-19: Artificial Analysis's model index returned no Doubao row, and Hacker News returned zero results across three queries for the model name and id. That is an absence of evidence, not evidence of weakness, and it is recorded as a limit rather than as a finding against the supplier. It has one practical consequence: a buyer's own fixed task set is the only measurement available, and a buyer who does not build one is buying on documentation alone.


The commercial surface is renminbi, Chinese-language and platform-bound. The rate card is published in renminbi with no dollar equivalent, the documentation set is published in Chinese, and the model is reached through a Volcano Engine account. A team outside mainland China has to settle eligibility, contracting entity, data location, payment and support language before it can compare this model on price with a model sold on a dollar card. Those are procurement facts rather than model facts, and they belong in the same decision as the rate.


One published choice runs in the vendor's favour and should be said as plainly as the rest. The price sheet publishes a single band across the whole 1,024K window, with no long-context multiplier, and it publishes cache storage at CNY 0.017 per million tokens per hour beside the cached-input rate. A vendor that could have charged a long-context premium and did not has made a reader's arithmetic simpler, and a vendor that publishes the storage line separately has made the cached-input rate checkable. Both are the kind of disclosure a buyer should notice.


No sub-score changed for these findings, and the reason is explicit. The Time sub-score already carries the setup overhead and the missing latency figure; the Quantity sub-score already carries the capacity and the route ratio; the Quality sub-score already carries the absence of measurement, which is the whole reason it sits at 4; and the Imposture assessment already carries the fluency risk of a very large window. The items above are about identity, evidence and procurement. None of them is a resolved failure and none is a defect in the artefact a buyer receives. Section 7c closes with what it does not do: it does not allege misconduct, it does not treat an unpublished benchmark as a hidden defect, and it does not convert a procurement obstacle into a capability judgment.


Back to the TOC


Section icon: U365 Co-Intelligence Rating.

Doubao Seed 2.1 Pro

U365 Co-Intelligence Rating


CI-First Profile


Primary profile: Co-Worker and Assistant (level 2). Secondary profiles: Co-Researcher, Analyst and Tester (level 4), Professor, Coach and Tutor (level 3).


Why this profile and not Co-Creator. The product is sold as a model that takes a large input and returns a finished answer: a document set reduced to a table, a repository slice reduced to a diff, an archive reduced to records. That is delegation, not collaboration on the thinking. The reader states the outcome, supplies the set, and reviews what returns. Co-Worker and Assistant is the role the model actually plays. Analyst and Tester is recorded as secondary because the long-context reading and the structured output are where the product is most useful and where the reader interprets rather than authors. Professor, Coach and Tutor remains reachable at level 3 because the thinking chain is visible and can be read as reasoning, and it is the weakest of the three readings: no teaching mode is claimed and the documentation is in Chinese.


CI-First Benefit Score


Time


  • Score (0-10). 6

  • Rationale. A structural saving on long-document work: a 1,024K input window and a 256K answer ceiling remove the chunk-and-reconcile loop, which is where the hours go in a contract set or a repository slice. Held at 6 because the saving is not measured, thinking is on by default, the answer default is 4K and has to be raised, and account setup plus Chinese-language documentation is real overhead.


Quantity


  • Score (0-10). 7

  • Rationale. A genuine multiplier on a workflow: the 256K answer ceiling, batch and low-priority routes at half the regular input and output rates, cached input at one fifth of the input rate, and structured output that turns a long response into records. Held at 7 because no throughput is published, the 1,000,000 tokens-per-minute limit is labelled non-rigid, and thinking tokens consume the output budget.


Quality


  • Score (0-10). 4

  • Rationale. Measured absence. The model list publishes capability labels, not benchmark results, and no independent index row was found on 2026-11-19. A quality claim without a measurement is not a quality benefit, so the architecture is not priced as a result. Four is the framework's marginal position, and it is the honest one until a third party measures the model or the reader measures it on a fixed task set.


Skill


  • Score (0-10). 3

  • Rationale. Delegation rather than learning in the common case, with no teaching mechanism, no memory and Chinese-first documentation, scored conservatively per the framework's guidance. The visible thinking chain is the one learning surface, and reading it is a habit the model does not teach.


CI-First Benefit Score: 5.0 / 10 (CI-First Positive)


Why the score sits where it sits


5.0 with sub-scores 6 / 7 / 4 / 3 is a capacity-and-volume position carrying a narrow time position, an unmeasured quality position and a marginal skill position, and every part of that sentence is deliberate.


Why not higher. Two reasons, both evidential. There is no independent measurement: no index row, no published head-to-head, no review-platform record. And the quality dimension is where the framework refuses to price an architecture as a result. A model with a 1,024K window and a 256K answer ceiling is a real capability event; it is not a measured quality event, and the score separates the two.


Why not lower. The capacity is genuinely useful and the price shape is genuinely generous: one band across the whole window, batch and low priority at half rate, cached input at one fifth, and a published cache-storage line. For a document-heavy workflow, that combination removes work rather than adding it. The band stays CI-First Positive, which is a real recommendation with disciplined use, and a reader should not read 5.0 as a rejection.


Why 5.0 and not a different band, stated as arithmetic. Time 6, Quantity 7, Quality 4 and Skill 3 average to 5.0. Under framework v1.4.1 the band is a function of the score: 4.1 to 6.0 is CI-First Positive. The band is stated from the score rather than set beside it.


What the score does not carry. The four dimensions measure what the reader's work gains. They do not carry the procurement questions, the renminbi-only rate card or the two-snapshot identity problem, 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 vendor risk would need a different instrument; the framework's instrument is the clause note and the 7c finding.


Humics Protection Badge


Creativity


  • Rating. Neutral (0)

  • Rationale. It produces tables, records, diffs and summaries on request, and it does not originate the direction. A million-token window makes more material available to a reader's own thinking, which is a small protection, and the model neither trains nor replaces the reader's ideation. The judgement about what matters stays with the reader.


Critical Thinking


  • Rating. Erodes (-1)

  • Rationale. Long, fluent analysis of a document set whose correctness is expensive to check, produced with no independent measurement to calibrate trust and with a window that invites handing over the whole set. The failure mode is not an obvious error; it is a comparison table with a plausible cell in every row and a record set with no nulls. The checklists in Real Workflows lower the frequency and do not reverse the direction.


Social Authenticity


  • Rating. Neutral (0)

  • Rationale. The model produces text and structured data 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.


Humics Protection Score: -1 / +3 Badge: Humics-Neutral


Superhuman Usage Guidance


When to invite this tool:


  • Whole-set reading where the alternative is chunking: a contract set, a filing set, a reporting pack, a codebase slice.

  • Volume extraction against a stable schema, on the batch route, where nobody needs the answer today.

  • Work where a qualified reader will check the output and the citation is required in every cell.

  • Teams already contracted on Volcano Engine, where account, region and payment are settled and the model is one more endpoint.

  • The task where the set is too large for a 128K model and the answer needs to be longer than a summary.


When to keep this tool out:


  • Short requests, where thinking dominates the cost and the wait.

  • Any output that will be acted on without a qualified human reading it. The longer and more complete the answer, the more convincing the error.

  • Procurement that cannot contract with a Chinese cloud platform, or a budget that must be stated in dollars at a named rate.

  • Latency-sensitive workloads, because there is no low-latency tier for this model.

  • Speech and audio workflows, because audio input is not priced for it.

  • A decision that requires an independently measured quality claim before adoption. That evidence does not exist yet.


U365 method integration:


  • LIPS + CARE: route the model's reading and extraction output into the Collect phase, and keep the Action Plan and Review phases as your own work. The model can produce the comparison and the records; it cannot own the decision about what they mean.

  • ULM + EVA: the decision about which work is worth automating belongs in the Career and Finance domain and in Quality of Life planning, and cost per completed task is the arithmetic that decision needs. The route choice, regular, low priority or batch, is a planning habit that sits with the Fellow.

  • UP-Context-Engineering: the method's eight layers are the control surface this model needs, because a stated AI-EXACT-TASK and an EXPECTED-FORMAT-AND-CONSTRAINTS that carries the limits are what keep a million-token request inside the job it was given. FULL-CONTEXT is where the window is spent, and spending it deliberately is the skill.

  • UNOP: no direct fit. The model is not built to teach, and its gains come from capacity and disciplined execution rather than from any pedagogical mechanism.


Over-delegation warning: the failure mode with this model is that capacity reads as permission. A window that holds everything makes it easy to stop deciding what matters, and an answer that is long and well formatted makes it easy to stop checking. 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 a document set into the window because it fits, and changes nothing about how much of the output a qualified person reads, has taken the product's worst offer.


Back to the TOC


Section icon: What Users Say.

Doubao Seed 2.1 Pro

What Users Say


Aggregate Rating Table


Platform

Rating

Number of reviews

Link

Artificial Analysis

No Doubao row on the model index when checked on 2026-11-19

Independent benchmark, not a user rating

Hacker News, via the Algolia search API

Zero results for "doubao seed 2.1", "doubao-seed-2-1-pro" and "doubao 2.1 pro"

Not a rating platform

G2

No listing for the model

Not applicable

Capterra

No listing for the model

Not applicable

Trustpilot

No listing for the model

Not applicable

Product Hunt

No launch page for the model

Not applicable

Reddit

No thread discussing this model was found in the searches run for this review

Not a rating platform

Volcano Engine documentation and price sheet

Vendor documentation, not a rating

Not applicable


What the public record contains


Almost nothing that a buyer can use as sentiment. The model is sold through a Chinese cloud platform, its documentation is in Chinese, and the English-language review economy that produces G2 pages, launch threads and benchmark commentary has not covered this model at the time of writing. The searches behind the table above were run on 2026-11-19 and are recorded with their results rather than summarised into a mood.


What is missing from it, and why that matters less than it looks


A reader should not read an empty review record as a negative signal. A model sold to enterprises through a regional cloud platform often has no public review footprint at all, because the buyers are procurement teams rather than individuals. What the absence changes is the confidence a reader can place in second-hand reports. What it does not change is the model's published capacity, which is stated on the vendor's own pages and is what this review scores.


Sentiment Summary


Neutral by absence. There is no measurable user sentiment to report, positive or negative, and this review does not manufacture one from vendor claims or from the consumer Doubao assistant, which is a different product on a different surface.


U365 Editorial Note


The consumer Doubao app is not this model. App-store ratings for a chat assistant, in Chinese or in translation, describe a different product with a different interface, different limits and a different price, and a reader who carries them across has made an unsupported inference. If sentiment matters to your decision, run a small trial of your own on your own documents: that produces a measurement, which is more than the public record currently offers.


Back to the TOC


Section icon: Comparison and Alternatives.

Doubao Seed 2.1 Pro

Comparison and Alternatives


The comparison is drawn from the same price sheet read on 2026-11-19, so every figure below is checkable against one page. Prices are renminbi per million tokens.


Model

Input

Cached input

Output

Input window

Choose it if

Doubao Seed 2.1 Pro

CNY 6.00

CNY 1.20

CNY 30.00

1,024K on the current snapshot, 256K on the earlier one

You need the largest window in the family and a long answer, and you will pin the snapshot

Doubao Seed 2.1 Turbo

CNY 3.00

CNY 0.60

CNY 15.00

256K

Your inputs are under 256K and you want half the token price, and you accept a smaller window

Doubao Seed 2.1 Lite

CNY 0.80

CNY 0.16

CNY 2.70

1,024K

Volume matters more than depth, and audio input is part of the workflow: this is the model in the family with a published audio input rate

Kimi K3.1 Pro

CNY 13.00

CNY 1.30

CNY 68.00

Not published on the price sheet

You have already established that the task needs it, because it is more than twice this model's input rate and more than twice its output rate

GLM-5.2

CNY 8.00

CNY 2.00

CNY 28.00

Not published on the price sheet

You want a comparable output rate from another vendor and a slightly higher input rate

DeepSeek V4 Pro (formal version)

CNY 9.00

CNY 0.30

CNY 27.00

Not published on the price sheet

Cached input dominates your bill: its cache-hit rate is the lowest of the group, at CNY 0.30


Two further comparisons are worth naming, and neither is on this price sheet. Gemini 4 Neon is the model U365 reviewed on 2026-11-19 at 6.3 / 10 (CI-First Strong), sold at $3.00 per million input tokens and $15.00 per million output tokens with a 1,048,576-token window, and it is the nearer choice for a team that needs an independently measured Western model and cannot contract in China. Claude Fable 5.2 and GPT-6.2 sit in the same Western frontier set. This review does not state their scores, because a score belongs to the review that produced it and those documents are their own authority.


How to choose between them, stated as three questions. First, does the set fit 256K or does it need the million? That decides between Turbo and Pro inside this family. Second, is the work interactive or can it wait? That decides between regular, low priority and batch, and it changes the bill by half without changing the model. Third, can your organisation contract with a Chinese cloud platform at all? That decides whether any of the models above is available to you, and it is a question no benchmark answers.


Back to the TOC


Section icon: Verdict and Next Steps.

Doubao Seed 2.1 Pro

Verdict and Next Steps


The verdict. Doubao Seed 2.1 Pro is a serious long-context instrument with an unusually simple price shape, sold on a platform that many readers outside China cannot use and supported by no independent measurement that this review could find. Its 5.0 / 10 (CI-First Positive) is a real recommendation for the work it is built for, on three conditions: pin the snapshot, set the thinking and answer limits deliberately, and keep a check you own in every workflow.


What would change the verdict. A first independent evaluation at a published setting would move the Quality sub-score and could move the band. A published release date and an English documentation set would remove procurement friction without touching a score. A change to the single-band price, the batch ratio or the cached-input rate would move the cost argument. Each of those is a re-check trigger below.


Next steps, in the order that saves the most time


  • Settle the account question first. If your organisation cannot contract on Volcano Engine, stop here: the model question is moot, and the comparison table above tells you which Western model to read next.

  • Pin `doubao-seed-2-1-pro-260915` in code and write the pin into your architecture note. The family name on the price sheet is not callable, and the earlier snapshot has a quarter of the window.

  • Take one real document set and measure it. Record the input tokens, the thinking setting, the route and the output tokens from the response, then multiply by the published rate. That is your unit cost, and it is the only measured number you will have.

  • Turn thinking off for extraction and classification, and leave it on for analysis. Write the rule down, because the default is on and the output rate carries the thinking tokens.

  • Build a twenty-item fixed task set of your own. With no independent benchmark available, your own set is the measurement that decides adoption, and it also gives you a baseline for the next snapshot.


A UP-Context prompt pack for the adoption decision


This sample prompt uses the UP-Context-Engineering Method and the CI-First doctrine. Read the method at https://www.university-365.com/post/the-up-method-university-365-prompting-the-new-gold-standard-for-prompt-engineering, the doctrine at https://www.university-365.com/post/discover-the-ci-first-approach, learn both in a U365 online program at https://www.university-365.com/programs, and request a U.Coach for orientation at https://www.university-365.com/ucoach.


USER-PERSONA: My UP-Context-Engineering USER-PERSONA file is the file for this layer, attached. It is my psychographic file concatenated with the six ULM life-domain files. The parts that matter here: I sign the adoption decision and I have to defend it to a procurement committee.
FULL-CONTEXT: The FULL-CONTEXT set for this task, added and not merged: BUSINESS-PERSONA for [the organisation], PROJECT-DESCRIPTION for [the workflow under consideration], the document set [sample or manifest], the current tool and its cost [figures], the volume per month [figures], the procurement constraints [list], and the two documents that decide the question [names]. If the volume or the current cost is missing, ask for it and stop.
AI-PROFILE-ROLE-EXPERTISE: Act as a Co-Researcher, Analyst & Tester, in the role of an analyst comparing a hosted model against an incumbent workflow, with expertise in cost per completed task, procurement risk and evidence quality.
USER-FINAL-GOAL: A one-page decision note I can defend: the measured cost per task on the new model, the comparison with the incumbent, the procurement facts that gate it, and the evidence that is missing.
AI-EXACT-TASK: Compute cost per completed task for the new model on the supplied volume and give a recommendation with the condition attached.
AUDIENCE: A procurement committee that will not read the vendor documentation and will ask what we measured ourselves.
EXPECTED-FORMAT-AND-CONSTRAINTS: Format: (1) the cost arithmetic, (2) the comparison, (3) the procurement gates, (4) what we have not measured, (5) the recommendation with its condition. Constraints: state every assumption; name the evidence that does not exist; do not convert renminbi to dollars without naming the rate and the date; do not present a vendor claim as a measurement; do not recommend adoption on capability labels alone; if a figure is missing, write Missing rather than estimating it; keep the note to one page. Data limits: use only the figures I supply and the public rate card; no customer data in this prompt.
EXAMPLES: The last procurement note in this house style is attached as the format example.

Before you use the output, run the check yourself: re-derive the cost per completed task from the rate card and the token counts you supplied, confirm that every assumption is named, and confirm that the missing evidence is stated as missing rather than filled. The model's capacity is published and the price is published; the quality, at the time of writing, is not measured by anyone you can cite. The decision is yours, per CI-First.


Related U365 content



Back to the TOC


Section icon: Status and Last Tested.

Doubao Seed 2.1 Pro

Status and Last Tested


Status: Active | Last tested: 2026-11-19 | Version tested: Doubao Seed 2.1 Pro (`doubao-seed-2-1-pro-260915`, vendor snapshot token 260915; no release date published)


Active: the tool is current and recommended.


The model list, the price sheet, the product overview, the text-generation guide, the deep-thinking guide, the image-understanding guide, the GUI Agent guide and the model pricing page were read on 2026-11-19, and the two callable snapshot ids were confirmed in official code samples on those pages. 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 snapshot id, because the capacity of the product is expressed there and nowhere else.


Back to the TOC


Section icon: Migration Path.

Doubao Seed 2.1 Pro

Migration Path


Not applicable. Doubao Seed 2.1 Pro is Active, not Retired, Deprecated or Risky, so this review carries no migration plan away from the tool.


If you are moving a workload onto it, or from the earlier snapshot to the current one, the practical changes are short and all of them are configuration. The earlier snapshot doubao-seed-2-1-pro-260628 carries a 256K context window and a 256K maximum input; the current doubao-seed-2-1-pro-260915 carries 1,024K on both. The price row is the same, so the move is a capacity gain at an unchanged rate, and the things to re-test are the ones a larger window changes: prompt length, the position of the instruction inside a long input, the answer ceiling, and whether thinking is still worth its tokens on the longer set. A team moving from another vendor also changes the answer default, which is 4K here, and the thinking switch, which is thinking.type=disabled or reasoning.effort=minimal depending on the API surface. That is a configuration change rather than a migration away from a tool, and this review does not treat it as one.


Back to the TOC


Section icon: U365's Recommendations to Learn More.

Doubao Seed 2.1 Pro

U365's Recommendations to Learn More


Start with the model list, not with a launch article


The model list is the document that settles identity, capacity and rate limits in one place: the two callable ids, the 1,024K window, the 256K answer, the 256K thinking chain and the 500 RPM and 1,000,000 TPM limits. A reader who starts anywhere else will meet the family name and assume it is the model id. Read this page first and keep it beside your architecture note.


Then read the price sheet for its routes as well as its rate


The interesting facts on the price sheet are structural: one band across the whole window, batch and low priority at exactly half the regular input and output rates, cached input at one fifth of input, and a separate cache-storage line. A reader who takes only the headline rate has taken the most expensive route and the least useful fact.


Then the two guides that change how a call is written


The deep-thinking guide is where the thinking switch and the thinking budget are explained, and the image-understanding guide is where the multimodal surface is shown with this model's id in its own samples. Between them they decide the two settings that most change the bill and the answer.


What to read instead of a benchmark table


There is no benchmark table to read. The honest substitute is your own: twenty items, a fixed scoring rule, and the same twenty items again when the next snapshot appears. Write down the snapshot token beside the result, because a score without the snapshot is a score of a product that may no longer be the one you are calling.


Three videos worth the time


  • Standarity, "What is Doubao Seed 2.1?": a short English overview of the family for a reader meeting the name for the first time, and the fastest way to place Pro against Turbo and Lite. https://www.youtube.com/watch?v=ejY4MPMmwnI

  • 鲲鹏Talk, "Doubao Seed 2.1 Pro launches today": a hands-on run reporting 58 tokens per second with the model wired into ChatGLM and TRAE, in Chinese, and the closest public look at the Pro snapshot under load. https://www.youtube.com/watch?v=F03ULh_19Kc

  • 0xTuan, "Doubao Seed 2.1 Pro released": multimodal input and the published renminbi rates with a one-shot pixel-game test, in Chinese. It quotes the 256K window, which is the earlier snapshot rather than the current one. https://www.youtube.com/watch?v=ZYcmI6biCjw



Resources on X


Dedicated X channels: no English-language X channel covers this model, and the handle this review found under the Doubao name belongs to the consumer assistant rather than to the Pro snapshot, so the image below is recorded as a signpost and not as a measurement channel. The practitioner material for this model is published in Chinese on Chinese platforms, which is the same procurement fact the review records elsewhere.


The avatar of the account posting under the Doubao name on X, recorded so a reader recognises it rather than mistaking it for this model's channel

U365 courses and programmes that build the surrounding skill


  • UIT (Technology, AI, Data Science): the engineering track where context engineering, structured output and cost per completed task belong.

  • UIB (Business Management, Entrepreneurship): the commercial track where a rate card with three routes and a procurement gate are analysed.

  • UIC (Digital Communication, Marketing): the track where a long-context drafting surface meets a publication standard.


The one habit to carry


Write the snapshot token into the same line as the result. A model whose capacity is expressed in its id, sold on a price sheet that names the family, will otherwise change under your measurement without telling you.


Back to the TOC


Section icon: CI-First Evaluation Summary Card.

Doubao Seed 2.1 Pro

CI-First Evaluation Summary Card


Field

Value

Tool

Doubao Seed 2.1 Pro (doubao-seed-2-1-pro-260915), ByteDance, served on Volcano Engine

Category

Applied AI / Large Language Model (long-context document work, agentic tool use, multimodal input, China-hosted)

CI-First Benefit Score

5.0 / 10, CI-First Positive

Sub-scores

Time 6 / Quantity 7 / Quality 4 / Skill 3

CI-First Profile

Primary: Co-Worker and Assistant (level 2). Secondary: Co-Researcher, Analyst and Tester (level 4), 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 Medium, Quantity Medium, Skill Illusion Medium

Framework v1.2 clauses

5.2.3-a null with boundary, 4.2-a null with live condition

Clause 7.5

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

Status

Active

Last tested

2026-11-19

Version tested

Doubao Seed 2.1 Pro, doubao-seed-2-1-pro-260915, vendor snapshot token 260915, no release date published

Pricing

CNY 6.00 per million input tokens, CNY 1.20 cached input, CNY 30.00 per million output tokens, one band across the 1,024K window; batch and low priority at half rate

Primary institute

UIT (Technology, AI, Data Science)

Review status

Active: the tool is current and recommended


Back to the TOC


Section icon: Glossary.

Doubao Seed 2.1 Pro

Glossary


CI-First Benefit Score


The arithmetic mean of four benefit dimensions, each scored 0 to 10, rounded to one decimal place. It answers one question: does this tool make Artificial Intelligence more profitable than Human Intelligence alone, net of the overhead of using it? The band is a function of the score: 0 to 2.0 is CI-First Negative, 2.1 to 4.0 Neutral, 4.1 to 6.0 Positive, 6.1 to 8.0 Strong, and 8.1 to 10.0 Transformative.


Time Benefit


Net time saved after prompting, configuring, reading, verifying and correcting. A tool that saves an hour of writing and costs fifty minutes of checking scores on the ten minutes, not on the hour.


Quantity Benefit


How much more usable, verified output a reader produces in the same time. Volume that does not survive verification does not count, which is why this dimension is read beside Quality rather than instead of it.


Quality Benefit


Whether the output is better than the reader could produce alone, judged after verification rather than on surface polish. A quality claim without a measurement is not a quality benefit, and the absence of measurement is scored as an absence.


Knowledge and Skill Benefit


Whether the reader learns something durable, or only receives an output. This is the dimension most exposed to the Skill Illusion, and the framework's instruction is to score it conservatively.


Humics Protection Badge


The sum of three ratings, each +1, 0 or -1, for Creativity, Critical Thinking and Social Authenticity. Plus two to plus three is Humics-Friendly, minus one to plus one is Humics-Neutral, and minus two to minus three is Humics-Risky.


AI Imposture Risk


The likelihood that a tool traps a reader in one of three illusions: the Time Illusion, where the appearance of speed hides overhead; the Quantity Illusion, where volume looks like quality; and the Skill Illusion, where the appearance of competence hides a missing capability. Each trap is rated Low, Medium or High, and the overall level follows the framework's arithmetic: all three Low is Low, one to three Medium is Medium, and two or more High, or a core value proposition built on an illusion, is High. Where two ratings could apply to the same trap set, the higher rating governs.


User Sentiment


The direction of independent and practitioner reaction to the tool, read from what testers and users actually reported rather than from a rating average, because no review platform rates this model at all. For Doubao Seed 2.1 Pro the record is empty on the English-language platforms, and the Chinese practitioner material describes a capable long-context model sold at a renminbi rate card behind a platform account, which is a procurement fact rather than a sentiment.


CI-First Profile


The role the AI plays in the working relationship, from a fixed set of five: Co-Creator and Thought Partner, Co-Worker and Assistant, Professor, Coach and Tutor, Co-Researcher, Analyst and Tester, and Challenger, Critic and Devil's Advocate. A review names a primary profile and any secondary profiles.


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 2026-11-19 and is re-checked on a trigger list with a six-month outer limit, because a hosted model's ids, limits and price routes move faster than a calendar interval can track.


Back to the TOC


Section icon: Sources.

Doubao Seed 2.1 Pro

Sources


Vendor primary sources, all read on 2026-11-19:



Independent measurement and community, checked on 2026-11-19:



U365 internal reference:


U365 CI-First Evaluation Framework v1.4.1 (URC methodology)


Back to the TOC


Section icon: Faculty Note on Evidence Quality.

Doubao Seed 2.1 Pro

Faculty Note on Evidence Quality


What this review rests on, in order of weight. First, the vendor's own primary documents: the model list, the price sheet, the product overview and the four capability guides, all read on 2026-11-19, with the callable ids confirmed in official code samples rather than inferred from a page title. Second, the framework's application rules for a model with no independent measurement. Third, the absence itself, which is recorded with the date and the sources checked rather than described as a general impression. No number in this review is attributed to a page whose content could not be read.


The vendor's numbers and the independent numbers do not meet, because there are no independent numbers. This is the central evidential fact of the review. The vendor publishes limits, capabilities and prices, and does not publish a benchmark table on the pages read. Artificial Analysis returned no Doubao row, and Hacker News returned no discussion across three queries. The Quality sub-score of 4 is the framework's response to that position: marginal, unmeasured, and stated as such in the rationale rather than softened by the model's architecture.


The comparison base is the recurring problem in this release's public surface. Two facts are easy to read wrongly and both depend on a base. The family name doubao-seed-2.1-pro is a billing row, not a callable id, so a cost estimate built from the price sheet alone can describe a 256K product at a 1,024K price. And a 1,024K window is a capacity, not a speed or a quality: a reader who reads a million tokens as a promise about latency or accuracy has moved from a published limit to an unpublished claim. Neither sentence is a lie. Both are sentences whose meaning depends on a base, and the framework's discipline is to state the base.


What is measured and what is inferred. The limits, the capabilities, the rate limits, the prices and the routes are published by the vendor and quoted here with their page. The claim that this model is the right instrument for a document set that no longer fits a 128K window is an interpretation of those limits 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 window size and the fluency of long output 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 price sheet publishes a single band across the whole 1,024K window with no long-context multiplier, and it publishes cache storage at CNY 0.017 per million tokens per hour as a separate line rather than folding it into the cached-input rate. A vendor that could have charged a long-context premium and did not, and that makes the cache arithmetic checkable, has done the reader a service. The model list also publishes both snapshots side by side, including the older 256K one, which is what makes the identity finding in Section 7c checkable at all.


What this review does not do. It does not score the consumer Doubao app, the Doubao Assistant, Seedance or Seedream, none of which is this product. It does not infer a release date from a snapshot token. It does not convert renminbi to dollars, because a converted figure without a named rate and date is a number a reader cannot check, and this review would rather publish no figure. It does not treat the absence of a benchmark as evidence of weakness. And it does not convert a published limit into a promise: a model with a 1,024K window will still return a comparison table that a qualified reader has to check, and the reader who plans for that is the reader the framework is written for.


Section icon: Status and Re-check.

Doubao Seed 2.1 Pro

Status and Re-check


Status: Active | Last tested: 2026-11-19 (Doubao Seed 2.1 Pro, `doubao-seed-2-1-pro-260915`, vendor snapshot token 260915) | 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:


  • A new snapshot id, or the withdrawal of `doubao-seed-2-1-pro-260915` or `doubao-seed-2-1-pro-260628`. The capacity of the product is expressed in the id, so a new id is a new product for the purpose of this review.

  • A change to the window, the maximum answer, the default answer length or the thinking-chain ceiling on the current snapshot. Time and Quantity rest on those four numbers.

  • A change to the renminbi rate card, to the single-band rule, or to the batch and low-priority ratio. The cost argument in Comparison and Alternatives depends on the 2:1 ratio and on the absence of a long-context multiplier.

  • The first independent evaluation of the model at a published setting. This is the trigger that would move the Quality sub-score, and it is the most important one on the list.

  • A published release date, an English documentation set or a non-Chinese region. Any of the three removes procurement friction without touching a score, and each changes what a reader can plan.

  • A change to the thinking defaults or to the two switches. Thinking on by default is a cost decision, and a default change moves every bill in the review.

  • A low-latency tier, an audio input price or a long-context multiplier appearing for this model. Each would add or remove a route the review currently records as absent.

  • A documented mechanism by which an agent wrapper writes or revises a reader's memory, skills or standing instructions on this model. That is the boundary in clause 5.2.3-a, and crossing it changes the clause outcome.


Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
Image by Erik  Lucatero

Become Superhuman

Master AI to stay irreplaceable in every field.

 

 

 

​

​

Apply for Admission Today.
Select Your Initial Access Level.


Become a DISCOVERY, INSIDER, or SUPERHUMAN Fellow.

Image by Milad Fakurian

Master Your Life with a Digital Second Brain

Turn overwhelm into clarity with LIPS + CARE
U365’s unique framework to organize your goals, projects, and knowledge into a superhuman system for success

bottom of page