Genspark: agents, an office suite and a memory layer on one credit meter
Status: Active | Last tested: 2026-09-28 (genspark.ai, its help centre, its terms, its privacy policy and its business pages, as they described the product on that date) | Re-check: trigger-based (max 6 months)
Active: the tool is current and recommended.
This review covers Genspark as one product: the web workspace, the agent surfaces inside it, the desktop and mobile applications, and the team and enterprise editions, as the vendor documents them in September 2026. The scope is Genspark alone.
Genspark scores 4.8 out of 10 on the U365 CI-First Review, which is CI-First Positive, with a Humics-Risky protection badge and a Medium AI Imposture Risk carrying Skill Illusion High. One subscription carries an agent engine, an office suite, a memory layer and a team room, and the meter that connects them is the part of the product a buyer has to reason about.
For detailed explanations of the CI-First evaluation terms used in this review, including the CI-First Benefit Score, the CI-First Profile, the Humics Protection Badge, the AI Imposture Risk levels and User Sentiment, see the Glossary at the end of this post.
In this Tool Review
Tool Snapshot
Genspark (Genspark AI Workspace)
Tagline: "Genspark - Your All-in-One AI Workspace" (genspark.ai, read 2026-09-28.)
Category: A single subscription that bundles an autonomous task agent, an AI-native office suite, a content suite, a personal memory layer, a persistent assistant that runs on a cloud computer, and a team workspace in which named AI agents hold roles alongside people. The vendor describes the current generation as Genspark AI Workspace 6.0.
Primary use cases:
Research and synthesis. A question becomes a structured, sourced research page, a briefing, a fact check or a comparison.
Work deliverables. Slide decks, spreadsheets, documents and dashboards, generated from a description or from source material you supply, then edited as normal files.
Recurring personal and team operations. Inbox triage and drafting, meeting notes and follow-ups, calendar work, scheduled reports and monitoring, spread across a workspace in which agents and colleagues share channels.
Media production. Images, video, audio and music from the same credit balance, from text prompts or from material you supply.
Agent-assisted building. Design and prototyping, code, and internal tools generated from a description.
What it is not. It is not a single model, and it publishes no benchmark of its own agent reliability. It is not a chat product with a task button; the agent surfaces are the product. It is not a self-hosted system, and it is not sold without a meter: almost everything above draws on a credit balance whose cost per run the vendor does not publish.
Pricing summary (read 2026-09-28):
Plan | Monthly billing | Annual billing | Credits | Storage |
Free | 0 dollars | 0 dollars | About 100 credits a day | 1 GB |
Plus | 24.99 dollars a month | 19.99 dollars a month, billed yearly | From 10,000 a month | 50 GB |
Pro | 249.99 dollars a month | 199.99 dollars a month, billed yearly | From 125,000 a month | 1 TB |
Team | 30 dollars per seat a month | Not published | 10,000 credits per seat a month on the business pricing card; 12,000 on the help centre page for the same plan | 60 GB per seat |
Enterprise | Custom, typically six figures annually | Custom | 20,000 credits per seat a month on the business page; 25,000 on the help centre page | Negotiated |
Two pricing surfaces disagree about the same paid plans, both read on 2026-09-28: the business pricing card prints 10,000 credits per seat for Team and 20,000 for Enterprise, while the help centre prints 12,000 and 25,000. The pricing page itself redirects to a sign-in, so the rate card is not readable without an account. Treat the printed tiers as the entry point rather than as a fixed price: Plus and Pro both offer higher credit tiers at higher prices, and the in-product plan page is the only place the full ladder is visible.
Credit costs, as the vendor and third-party trackers describe them:
Super Agent, the headline capability, carries no published per-run price. The vendor states that consumption varies with conversation length, tool use and model complexity.
A research page is estimated by third-party trackers at 5 to 80 credits depending on depth; a full slide deck at 100 or more; a phone call at roughly one credit a second, or about 180 credits for three minutes; video generation is the most expensive task class.
Core chat and core image generation are included at no credit cost, subject to the fair-use window described below.
Credits do not roll over. A downgrade resets the balance rather than carrying it, and an upgrade that changes the billing interval revokes unused credits and issues a fresh set.
Free quota and fair-use position (read 2026-09-28): Plus and Pro include a weekly free quota covering standard-mode text and image tasks across the agent surfaces, with the seven-day window starting at your first message. Where a task is covered, the composer shows a notice that the task is running at no credit cost. When the quota is exhausted, tasks continue on your credit balance unless you switch on the confirmation option on the usage page. Users report a separate session limit that resets on a fixed five-hour window; the vendor confirmed a session-limit change publicly in April 2026.
Platforms and access:
Surface | Where it runs | Notes |
Web workspace | genspark.ai | The full product, including the agent surfaces and the suites |
Desktop applications | macOS, Windows, Linux, plus dedicated apps for the persistent agent | The desktop app carries the local-execution mode of the persistent assistant |
Mobile applications | iOS and Android (workspace, team chat, mail, dictation) | The mobile listing shows more than one million downloads |
Office plug-ins | PowerPoint, Excel, Word; a Google Workspace add-on for Docs, Sheets and Slides | Work inside the file you already have open |
Hardware | A card-thin voice recorder sold with the memory layer | Sold separately, with a recording-consent guide published in the help centre |
Team and enterprise | Admin console, single sign-on, provisioning, governance controls | Seat-based; the enterprise tier adds residency and a support commitment |
Official links:
Website: https://www.genspark.ai
Help centre: https://www.genspark.ai/helpcenter
Terms of service: https://www.genspark.ai/terms
Privacy policy: https://www.genspark.ai/privacy
Business pages: https://www.genspark.ai/business
Device store: https://shop.genspark.ai
Trust page: https://trust.genspark.ai
No public status page was found at a dedicated address on 2026-09-28.
Underlying technology, as the vendor states it: A multi-model architecture in which each task is routed to whichever frontier model the system judges best suited, drawn from several model providers and from open-weight releases, coordinated by an agent engine the vendor describes as a mixture of agents over more than 80 integrated tools. The vendor's own materials describe four layers in the current generation: a memory layer, the agent engine, the suites where work is produced, and a collaboration layer in which people and agents share rooms. Where third-party reporting has described the architecture in more general terms, it repeats the same description. The system runs in the vendor's cloud for the agent surfaces, with a local-execution option for the persistent assistant's desktop mode and a connector set that includes the Model Context Protocol.
Company and legal entity: The contracting party in the terms is MainFunc Inc., together with its subsidiaries and affiliated entities, including Genspark Inc. as a wholly owned subsidiary. The company is headquartered in Palo Alto, California, and its Google Play developer record lists a Wilmington, Delaware address. The company describes itself as founded by former employees of large platform companies and reports a Series B financing total of 485 million dollars at a valuation of 2.6 billion dollars as of June 2026. It states a global partnership with Microsoft announced in April 2026 covering Office plug-ins, cloud deployment and availability through Microsoft's agent marketplace. The business page prints "Trusted by 7,000+ world-class teams"; an agency report in June 2026 put the count of business clients at more than 6,000 within six months of the business edition's launch.
Community evidence at a glance (each read on 2026-09-28):
Surface | Rating | Volume | What it covers |
Google Play, Genspark AI Workspace | 4.5 out of 5 | About 101,000 reviews, more than 1 million downloads | The mobile application, in a listing that mixes long-time users with billing complaints |
Apple App Store, Genspark AI Workspace (United States) | 4.71 out of 5 | 4,244 ratings | The mobile application |
Trustpilot, genspark.ai | 3.8 out of 5 | 443 reviews, 434 of them in the last 12 months | The company and the service, on a claimed profile with a paid subscription |
G2 | 3.7 to 3.8 out of 5 | Five to seven reviews, depending on which third-party comparison read the listing | Too thin to weigh, and read here as a qualitative signal only |
Product Hunt, per-launch | 4.0 out of 5 on the browser launch; 4.45 out of 5 on the Word plug-in launch | 11 reviews each | Individual launches, not the workspace |
Read the figures as four different instruments rather than as an average. The app-store numbers describe a large base of mostly satisfied mobile users. Trustpilot describes a smaller, angrier paying cohort, and it has moved sharply: earlier in 2026 the same profile sat near 1.5 out of 5 with a very high share of one-star reviews, and community threads from that period quote the profile at 1.7 out of 5 with 79 percent one-star. The current 3.8 is a recovery on a much larger review count, not a stable record.
Inputs: Text instructions, files up to 500 MB on the free plan and 1 GB per file on paid plans, connected data sources (mail, calendar, chat, documents, CRM), and a library of pre-built agent roles and scheduled tasks. The persistent assistant also accepts input from messaging channels and by email.
Outputs: Research pages with sources, slide decks, spreadsheets, documents, dashboards and internal tools, images, video, audio and music, drafted email, meeting notes and follow-ups, scheduled reports, and files stored in the account's own drive. Deliverables are downloadable, and some surfaces can be exported into the file types the office suites use.
Documented limits and stated positions worth knowing before you subscribe:
Session limits on chat and images apply alongside the credit balance, on a window that users report as five hours.
Failed runs can consume credits; multiple user reports and at least one formal review describe credit consumption on tasks that errored rather than delivered.
The cloud computer that hosts the persistent assistant is wiped when a cancellation takes effect, and the vendor states plainly that the data on it cannot be recovered.
The free plan's daily allowance is described as a lifetime free-credit allowance on the help pages, so it is not an indefinite entitlement.
Commercial-use rights are stated for paid members, with one vendor page guaranteeing them through 30 June 2027 and another stating 31 December 2026 for team plans.
One-line summary: A genuinely broad workspace that turns instructions into finished work across research, documents, media and operations, sold on a meter you cannot fully price in advance, with no published measure of whether the work it returns is correct, and with the strongest documented record in this series for two things at once: scale of adoption and severity of billing complaints.
The Problem
The promise of an AI workspace is that the software does the work rather than describes it. The problem it addresses is real and specific: the distance between having material and having a finished deliverable. A researcher has forty sources and no report. A marketer has product notes and no campaign. An analyst has a spreadsheet and no deck. A team has a decision in a chat thread and no record of it anywhere.
That distance used to be closed with a stack of subscriptions and a person who knew how to use each one. The promise of a single workspace is that the distance closes in one place, on one bill, with the work coming back finished enough to edit rather than finished enough to throw away.
Two failures follow the category, and Genspark meets both.
The work arrives without a measurement. An agent that returns a complete deck, a researched briefing or a drafted reply is returning an artefact you are asked to judge. The vendor publishes no task-completion rate, no hallucination rate and no accuracy figure for the agent surfaces. The one performance number in circulation, a benchmark score the vendor claimed in 2025, is a self-report that does not appear on the official leaderboard for that benchmark. Independent commentary says the same thing in plainer terms: there is no published measure of whether the agent completes the task correctly, so the reader is the verification layer, and the reader is the person least equipped to be one because the output looks finished.
The work is metered in a unit you cannot price before you run it. The credit is the billing unit, and for the product's headline capability the vendor publishes no per-run figure. You learn what a run cost after it ran. That is a defensible design for an agent that decides its own tool calls, and it is also the structural reason a user cannot budget the tool. The billing complaints in the public record are not about the price level; they are about not being able to predict or see it.
What happens without a tool like this. The work still gets done, by a person, at the speed of a person, in five applications that do not know about each other. Nothing in that older method fails. It is simply slower, and it does not benefit from the two things this category genuinely provides when it works: parallel execution of mechanical steps, and a memory of what you already asked for.
The Outcome
What a user actually gets from Genspark:
Finished artefacts rather than drafts. A described deck arrives as a deck, a spreadsheet arrives as a workbook with working formulas, and a research question arrives as a structured page with cited sources. The output is editable as a normal file rather than trapped in a chat log.
Parallel and unattended execution. The agent decomposes a task and runs the parts without supervision, and scheduled tasks and background check-ins run on a clock rather than on a request. Work continues while you do something else, which is the dimension where a workspace beats a chatbot by the widest margin.
One balance across a wide range of work. Research, decks, sheets, documents, email, images, video, audio and code draw on the same account and the same credit pool, which removes the coordination cost of five subscriptions.
A memory layer that removes the setup cost of every session. Once sources are connected, the assistant holds context about your work across sessions and across the channels you use, which is the difference between a tool you brief and a tool that already knows.
A team surface where the agents are named and visible. Agents hold roles, work in shared rooms, file tasks and deliver files back into the conversation, which makes the delegation boundary inspectable rather than invisible.
Enterprise-grade controls at the tiers that pay for them. Single sign-on, provisioning, role-based access, audit logs, model governance, residency options and a support commitment, with a certification position the vendor publishes.
What a user does not get. No measurement of whether the agent's work is correct. No published per-run cost for the headline capability. No write-approval step on the memory layer. No independent audit of the reliability claims. And no guarantee that the promotional pricing, the free quota or the unlimited arrangements survive the end of the promotional period, because they are stated as arrangements with end dates rather than as permanent terms.
For U365 the outcome sits in three places. A Fellow in any institute can use the workspace to produce the artefacts their programme assesses: a researched briefing, a structured deck, a workbook, a set of campaign assets. A team can use the room surface as a worked example of delegation with named boundaries, which is the Centaur discipline taught in the CI-First framework. And the whole product, including its billing design and its memory design, is the clearest single object the U365 corpus now carries for teaching the two traps that matter most in an agentic tool: work that looks finished and cannot be judged, and a meter that prices after the fact.
Who Should Use Genspark
Use it if you want:
A single workspace that converts instructions into decks, sheets, documents, research pages and media without a toolchain.
Unattended execution: scheduled reports, monitored sources, a briefing waiting for you, a recurring follow-up drafted before you ask.
A memory layer that holds your working context across sessions and channels, so you stop re-briefing the machine.
A team surface where AI agents hold named roles beside colleagues, with the delegation boundary visible in a shared room.
An office-plug-in path if your institution is standardised on Microsoft 365 or Google Workspace.
Do not use it for:
Any deliverable whose correctness you cannot verify yourself. The tool publishes no accuracy measure, and the output is formatted to look finished.
Anything that turns on cost predictability. Credits do not roll over, the headline capability carries no published unit price, and failed runs can consume the balance.
Material that must not leave a third-party processor, unless the tier you are on carries the controls that make the transfer acceptable to your institution.
Consequential communication sent in your name without a human read. The drafting surfaces write in your voice, and the persistent assistant can send.
Work where the memory layer would hold context you have no routine practice of reading back.
U365 Fellow categories:
Fellow type | Fit | Why |
Explorer | Good | Immediate, visible results on ordinary tasks, and the free plan is a real if limited way to meet the product |
Creator | Strong | The widest artefact range in this series, from decks to video, with editing in normal file formats |
Builder | Medium | The design, code and internal-tool surfaces are real, but the platform is closed and the API path is not the point of the product |
Researcher | Useful, with the caveat | Strong synthesis and briefing surfaces, and no published reliability measure, so the researcher remains the verification layer |
Practitioner | Conditional | Valuable where the deliverable is internal and reviewable, weak where the cost and the output both need to be predictable |
When to invite it, and when to keep it out. Invite it when the deliverable is internal, reviewable, and cheap to redo if wrong: research briefings, first drafts, first-pass decks, recurring reports, asset batches. Keep it out of anything consequential that the tool would complete and you would not be able to check, and out of any workflow where the cost of a run has to be known before the run happens.
U365 Institutes Alignment
Genspark is general-purpose, so this section rates what each institute can take from a workspace that produces artefacts across every one of their domains, and then states the limit that holds each rating.
UIT (Technology, AI, Data Science)
Rating. Medium (primary)
Why. The competency that remains is evaluation, and it is the most transferable thing this product carries. An agent workspace that publishes no task-completion measure asks the Fellow to state what measurement they would demand before trusting one, and to judge the architecture they are buying into from what the vendor discloses rather than from what it withholds. The second reading is metering a probabilistic system: a credit balance under a variable-cost agent, with no published unit price for the agent itself, no rollover, and failed runs that can consume the balance, which the Fellow turns into a cost per finished artefact and a stated limit under which a run may proceed
The limit that holds the row. Nothing here is open to read. The models are routed rather than specified, the agent engine is a black box, no architecture paper is published, and no benchmark run with a stated method is published for the current product. A cohort studies it as a consumer of the technology rather than as a system they can inspect or build against
UIB (Business Management, Entrepreneurship)
Rating. Medium
Why. Two commercial judgements survive the removal of the tool, and both are read from the vendor's published record. Cost appraisal: a workspace that meters a dozen capabilities in one currency, declines to price its flagship capability per run, does not roll credits over and revokes unused credits on an interval change, so a Fellow computes a cost per finished deliverable from their own usage and defends a tier, a scope or a route against the number. Contract appraisal: one refund per account with windows that differ by plan and by region, a rejection clause for heavy use before the request, two surfaces that disagree on the same commercial-use date, and a training position stated for the business tiers and not for the individual plans, so deciding what has to be answered before a recurring charge is accepted is a reading exercise the Fellow owns
The limit that holds the row. The product teaches no management, finance or entrepreneurship content of its own, and neither judgement has a published assessment home: a word-start, accent-folded search over the title and the full description text of all 79 published programmes returned zero matches for cost, pricing, price, metered, billing, charge, refund, renewal, cancellation, unit economics, rate card, invoice, procurement, supplier, spend cap, usage-based, consumption, licence, license, copyright and intellectual property. The nearest near-hit is budget at one programme, the budgets module inside Project Manager Mastery (25 days, published), which is a project budget and not a metered AI balance. The row stands at Medium on the two judgements and not above it, because the product teaches no management content and publishes no case data that would let a cohort evaluate the tier design against outcomes
UIC (Digital Communication, Marketing)
Rating. Medium
Why. The competency that remains is the publication rule for agent-authored communication: what may leave an account without a person reading it, which voice it carries, and what the recipient is owed when the text was drafted by a machine and sent under a person's name. On this product that decision is live rather than theoretical, because the drafting surfaces write in the user's voice and the persistent assistant can send, and the vendor leaves the professional question of presenting agent-authored text as your own entirely to the user. The category's central case is available to study here as well: audience contact at machine scale, and what an automated reply does to a relationship
The limit that holds the row. The tool performs the craft. It does not teach message design, audience analysis, or register; it produces artefacts that a competent communicator would critique, and it publishes no standard a cohort could use to critique them. The teaching value is in the review discipline the product makes necessary, not in what it produces
UID (Digital Design, UX/UI)
Rating. Low to Medium
Why. Coursework observation only. Two usable readings. The design surface turns a described interface, poster or prototype into a rendered artefact, which gives a cohort something concrete to critique against a specification. The second is the appraisal discipline: a delivered composite or a generated dashboard has to be read for hierarchy, consistency and whether it holds at the size it will be used
The limit that holds the row. The tool performs the composition. There is no layer control, no grid, no token surface and no design-system export, and the outputs are judged rather than authored. No U365 design competency is exercised by generating an artefact that the tool composed
U365 methods, not an institute (UNOP, ULM, LIPS, CARE and the UP-Context Method)
Rating. Applicable
Why. The methods layer is relevant in two specific ways. The memory layer is the closest commercial analogue to LIPS the series has reviewed: a compiled personal context that feeds every subsequent task, which makes the LIPS discipline of curating what enters the record directly transferable. The second is the CARE cycle, where the tool is strong at Collect and Execute and silent at Review, which is exactly the step the framework requires a human to own
The limit that holds the row. The workspace holds the memory but keeps no consent record, no disclosure decision and no register of what was generated and sent. A person who does not keep that record elsewhere has no record, and the platform's own position is that the account holder is the record keeper
The sentence that holds across all four rows. No U365 institute should treat Genspark as the instrument that decides whether a piece of work may be published or sent. The product removes production requirements and supplies no standard, critique or measurement for the judgement it replaces. The teaching value is in the disciplines it makes unavoidable and gives no help with: what to feed the memory layer, what to verify, what may be sent in your name, and what a run was allowed to cost.
Relevance is not a credential, and on this tool the two diverge sharply. A rating says a cohort has something to learn by reading or using the product. A credential says U365 assesses that competency and issues something for it. Genspark exercises competencies at every institute and is assessed by a credential at none, because the credentials U365 issues assess what a Fellow can do rather than what a tool can do for them.
Tool to Skill to Credential
No published U365 credential assesses the use of a general-purpose AI workspace. That is the finding, and it is a statement about this class of tool rather than a catalogue defect: a workspace removes the production requirement across many domains at once, and U365 credentials assess a competency a Fellow holds rather than a platform they operate. The Skill sub-score of 4 records the same thing from the scoring side.
Every programme named below was read from the published Online Programs catalogue on 2026-09-28: 86 records, of which 79 are PUBLISHED, each read with its title, its step count and, where the record publishes one, its duration.
Directing an agent through a multi-step deliverable: decomposing the goal, writing the instruction, reviewing the intermediate steps, and accepting or rejecting the result
U365 competency. Agentic workflow direction
Credential. Nothing published assesses agentic direction itself. The catalogue carries AI Agents and Workflows Automation with n8n - Certificate (published) and Automate your Work with n8n - Certificate (published), both of which teach building automations on a specific platform rather than directing a general agent; Superhuman Expert with AI - Diploma (published, 30 days), which teaches the context-engineering work of directing agents; and Superhuman@Work - Certificate (published), which publishes the deployment of agentic workflows and is the only published programme whose own text uses the term. None of the four publishes an assessment of direction as a competency. Adjacent anchors, not assessment homes
Institute. UIT (Technology, AI, Data Science), no credential mapped
Curating a personal memory layer: deciding what may enter it, what must be excluded, and how to read back what has accumulated
U365 competency. Personal knowledge curation
Credential. No published U365 programme assesses personal knowledge curation or the governance of an AI memory store. The nearest published anchors are Superhuman Expert with AI - Diploma (published, 30 days), the one programme whose own description builds the digital second brain from the LIPS and CARE method, and the Superhuman pathway certificates (published): Superhuman@Work, Superhuman@Life and Superhuman@Learn. The diploma teaches the method and publishes no assessment of curating a memory store or of governing one. Adjacent anchors, not assessment homes
Institute. U365 methods surface (LIPS), no credential mapped
Verifying machine-produced deliverables: judging a researched briefing, a deck or a workbook for the errors that survive formatting
U365 competency. Verification of AI-assisted work
Credential. No published U365 programme assesses verification of AI-assisted output as a competency in its own right. Data Analyst Expert (published, 84 days) and Business Analysis Professional (published, 60 days) both build the analytical judgement that verification depends on, and neither publishes an AI-verification outcome; Superhuman@Learn - Certificate (published) addresses information accuracy directly, as a learning protocol rather than as an assessment of a machine's output. Adjacent anchors, not assessment homes
Producing communication assets across text, image, video and audio from one brief, and keeping the register your own
U365 competency. Multi-format content production
Credential. The published content and media programmes carry the craft: AI Creator Professional (published, 30 days), Content Marketing Specialist (published, 30 days), Social Media Marketing Manager (published, 30 days) and Video Production Specialist (published). Each assesses the craft inside its own pipeline, and none publishes an AI-workspace direction outcome. Adjacent anchors, not assessment homes
Institute. UIC (Digital Communication, Marketing), no credential mapped
Costing and governing metered agent work: forecasting the consumption of a run, and setting the limits under which an agent may spend
U365 competency. Meter and governance discipline
Credential. No published U365 programme assesses the governance of AI consumption. The closest published anchors are AI Business Specialist (published, 18 days), which carries AI adoption inside a business pathway, and Project Manager Mastery (published, 25 days), which publishes a budgets module inside a project-management pathway and does not reach a metered AI balance. Adjacent anchors, not assessment homes
Institute. UIB (Business Management, Entrepreneurship), no credential mapped
Where relevance is present but uncredentialed, this review says so rather than leaving it silent. Five competencies are mapped above, every one of them names a verified published anchor and states what that anchor does not assess, and none is presented as an assessment home.
Access levels and what they decide. University 365 has three academic access levels: DISCOVERY, INSIDER and SUPERHUMAN. Specialised diplomas and certificates carry Basic, Foundation and Expert levels. DISCOVERY Fellows can enrol in Basic-level programmes only, INSIDER Fellows in Basic and Foundation programmes, and SUPERHUMAN Fellows in all of them. University degree programmes carry a single Expert level and are open to SUPERHUMAN Fellows only. No per-programme access level is asserted here, because the catalogue does not expose one, and no credit transfer is asserted between any two programmes. None of the chains above stacks into a degree, and no degree consequence is asserted for any of them, because each names a certificate or a diploma pathway rather than a degree-bearing sequence.
How Genspark Works
Inputs: A plain-language instruction, optionally with files (up to 500 MB on the free plan and 1 GB per file on paid plans), connected data sources (Gmail, Google Calendar, Outlook, Slack, Notion, HubSpot, Salesforce, Google Workspace, Microsoft 365, and more), a library of pre-built roles and scheduled tasks, and channels: chat inside the workspace, messaging platforms, and email.
Outputs: Research pages with cited sources, slide decks, spreadsheets, documents, live dashboards, internal tools and small applications, images, video, audio and music, drafted email and meeting follow-ups, scheduled reports, and stored files in the account's own drive. Most deliverables arrive as downloadable artefacts you then edit as ordinary files.
Underlying technology, described as the vendor describes it:
Model layer. A routing layer over several frontier models and open-weight releases, described by the vendor as a mixture of agents that assigns each task to the model judged best suited and cross-checks results. Third-party reporting repeats the same description, and one vendor-page count of models on the business plans is "70+"; an earlier product generation was described as nine specialised models over more than 80 tools.
Agent engine. A task agent that decomposes an instruction, selects tools, executes steps in sequence or in parallel, and returns finished deliverables, with a workspace that holds intermediate files for the life of a session. The vendor documents running several tasks at once inside one project and states that the agent keeps working after the laptop is closed.
Memory layer. A compiled personal context built from connected mail, calendar, chat, documents, meeting notes, uploaded files and the workspace's own project history, exposed to the agent as standing context. The vendor states the compiled data is accessible only to the account holder, that integrations require explicit authorisation, and that they can be disconnected at any time.
Persistent assistant. A long-running agent that connects to messaging channels, keeps a dedicated email address, holds memory across channels, runs scheduled tasks and periodic check-ins, and can execute on a dedicated cloud computer or locally through the desktop application. The vendor publishes a prompt-injection warning for this surface and a local-mode security note stating the workspace folder is soft guidance rather than a hard boundary.
Collaboration layer. A shared workspace where named agents hold roles and take assignments alongside people, file tasks into a shared board, and deliver files back into the conversation. Each agent runs in its own isolated sandbox according to the vendor, and the vendor states that admin accounts cannot read individual members' project content.
Suites. An office suite (slides, sheets, documents, mail), a build suite (design and prototyping, code, internal tools and dashboards), and a content suite (chat, image, video, music, audio, transcription).
Connectors. More than 200 connections to external systems, including the Model Context Protocol, which is the standard that lets an agent reach a tool endpoint without bespoke integration.
Integrations that matter for a U365 learner: the PowerPoint, Excel and Word plug-ins and the Google Workspace add-on, because they put the assistant inside files the reader already works in; the Microsoft 365 connector set, including mail, calendar, files and Teams as a channel; and the messaging channels, which is where the persistent assistant meets the reader without requiring a new habit.
Where the product is deliberately thin on public detail: the model roster is not published as a specification, no architecture documentation is published, no task-completion or accuracy figure is published for the agent surfaces, and no per-run cost figure is published for the agent itself. Those absences are findings in this review rather than oversights in it.
Getting Started with Genspark
Required accounts: A Genspark account, which is free and gives about 100 credits a day. A paid plan from 24.99 dollars a month removes the daily ceiling, adds credits, adds storage and unlocks the video and audio model set. The persistent-assistant surfaces carry a separate subscription if you want the cloud computer rather than local execution.
Installation: Optional. The workspace runs in a browser. The desktop applications exist for macOS, Windows and Linux, and mobile applications exist for iOS and Android. The office plug-ins install into PowerPoint, Excel and Word, and a Google Workspace add-on installs into Docs, Sheets and Slides.
First-time configuration:
Create an account at https://www.genspark.ai and open the workspace.
Run one task in the free plan before paying, so you see the interface and the output shape on your own material.
If you will use the memory layer, connect one source only (mail or calendar), read the authorisation screen, and decide from there whether the second source is acceptable.
If you will use the persistent assistant, choose local mode first. It runs on your own machine with your own account and needs no additional subscription, and it makes the data position inspectable before you commit to a cloud computer.
If you are on a team plan, complete single sign-on and role assignment before inviting anyone, not after. The admin controls are the reason the tier exists.
First 15 minutes checklist:
☐ Ask the agent for a short research page on a topic you already know well, then read the sources it cites. This is the only reliable calibration you can do in fifteen minutes, and it tells you what verification will cost you.:
☐ Ask for a five-slide deck on the same topic and edit one slide. The point is not the deck; it is finding out how much of it is worth keeping.:
☐ Open the usage page and note the credit balance before and after both tasks. Run both on the free plan if you can, so the first lesson about the meter costs nothing.:
☐ Export or download one artefact and open it outside the workspace. Confirm it is a real file, not a preview.:
☐ Decide, in writing, which of your own documents must never enter this account. Do that before you connect a second source.:
Result: After fifteen minutes you should hold three things: a researched page and a deck you can critique, a real number for what one task consumed, and a written boundary for what you will not connect. The third is the one that protects you later, because the memory layer is the surface that makes the product useful and the surface that carries the most risk.
Note on the pricing gate: the free plan is usable, not a demo, but the daily allowance is small against the agentic surfaces. The first serious workflow in this review is designed to run on the free plan for that reason.
Real Workflows
Workflow 1: The sourced briefing you can defend, on the free plan
Learner type: Any
CI-First benefit tags: Time / Quality
Connects to: LIPS (method surface)
Time estimate: 35 to 45 minutes including verification
What you do vs what the tool does:
Step | You do | The tool does |
1 | State the question, the audience, the length and the standard you will judge against | Decomposes the task and selects the research and synthesis tools |
2 | Nothing during the run | Gathers sources, reads them, drafts a structured page and cites what it used |
3 | Open every cited source and check it says what the briefing says it says | Nothing. This step is yours alone |
4 | Mark every claim you cannot trace to a cited source, and delete or rewrite it | Nothing |
5 | Save the verified version into your own record and note what you cut and why | Stores the artefact and the run history in the account |
Sample prompt:
Context: I am preparing an internal briefing on a question I will have to defend in a written exam and in conversation with people who know the subject better than I do. Role: AI as a research assistant. I own the judgement about what is true and the final wording. Profile: Act as an Analyst and Tester, the one this task needs. Test the material rather than asserting a position. Task: Produce a structured briefing with (1) the question restated in one sentence, (2) the three positions that have real support, with the strongest evidence for each, (3) the evidence that weakens each position, (4) the points where sources disagree and why, and (5) a list of what would change my mind. Constraints: Cite a working source for every factual claim. Name any claim you could not source instead of asserting it. Do not average conflicting sources into a single position; present the conflict. Output format: A headed outline of at most 1,200 words, ending with a section headed What I could not establish. Memory: This briefing belongs in my own record, because the tool holds no standing instruction of its own. I will file the verified version myself. UP-Context verification: I checked every cited source myself and removed every claim I could not trace to one. Data safety: this pack carries no personal data. I do not paste unpublished institutional material, client names, or anything under a confidentiality obligation into it.Verification checklist:
☐ Multi-Model Check: run the same question in a second assistant and compare the source lists, not the prose. Divergence in sources is the signal.
☐ External Source: open each cited source and confirm it says what the briefing claims. A citation that resolves to a page that does not carry the claim is the failure mode this step exists for.
☐ Human Review: give the briefing to someone who knows the subject and ask them what is missing, not whether it is good.
☐ CI-First Test: can you explain and defend the briefing without the tool? If not, the briefing is not ready.
Workflow 2: The recurring team briefing that runs while you work
Learner type: Professional
CI-First benefit tags: Time / Quantity
Connects to: Microsoft 365 Expert
Time estimate: 30 minutes to build, then minutes a week
What you do vs what the tool does:
Step | You do | The tool does |
1 | Write the standing brief: what to watch, who it is for, what a useful entry looks like, what to ignore | Offers a template and a scheduling interface, and asks the questions it needs answered |
2 | Set the boundary: which sources it may read, which it may not, and what it must never send | Connects the authorised sources and holds the configuration |
3 | Review the first three runs closely | Produces the scheduled output on the clock and delivers it where you said |
4 | Correct the brief when the output drifts, and record what you changed | Applies the change from the next run onward |
5 | Decide what of the output actually goes to other people, and send it yourself | Nothing. Delivery decisions stay with you |
Sample prompt:
Context: We run a weekly operating rhythm. I need a briefing that is useful rather than long, and I need to know what it could not see. Role: AI as a monitoring assistant. I decide what is shared and with whom. Profile: Act as a Co-Worker and Assistant, the one this task needs. Execute the routine work and hand it back for review. Task: Every Monday at 08:00, produce a briefing with (1) what changed in the sources I have authorised, (2) the three items that most need a decision and why, (3) what you could not access and therefore could not check, and (4) a one-line note on anything you are unsure about. Constraints: Read only the sources I have authorised. Never send anything to anyone. Flag uncertainty rather than smoothing it over. If a source is unavailable, say so rather than guessing. Output format: A short structured briefing, headed by the date and ending with a section headed What this briefing could not see. Memory: The briefing is filed in my own project record, because the tool keeps no standing instruction of its own and I need the history in one place I control. UP-Context verification: I read the first three briefings in full before letting the schedule run unattended, and I check one item per week since. Data safety: this pack carries no personal data. I do not connect mailboxes, drives or channels holding confidential institutional material, and I do not paste anything subject to a duty of confidence into the brief.Verification checklist:
☐ Multi-Model Check: for the first month, compare one briefing item a week against a manual check of the same source.
☐ External Source: open the underlying item for any decision the briefing places in front of you. Never decide from the summary.
☐ Human Review: a colleague reads the first month of briefings for whether they would have missed anything.
☐ CI-First Test: can you describe the week's situation without the briefing? If not, you are depending on it rather than using it.
Workflow 3: The delegation you can audit, with agents in a shared room
Learner type: Professional
CI-First benefit tags: Quantity / Skill
Connects to: CARE (method surface)
Time estimate: 45 minutes to set the room, then per assignment
What you do vs what the tool does:
Step | You do | The tool does |
1 | Write each agent's task boundary before the room opens: what it may do, what it must not, what it hands back | Holds the roles and the shared context |
2 | Decide the review gate: what must be true before any agent's output is used | Files tasks, tracks them in the room, returns deliverables to the conversation |
3 | Review each member's output against its own boundary, not against the room's output as a whole | Nothing. Attribution is your job, which is why step 1 exists |
4 | Record the decision and who owns it in your own notes | Keeps the thread and the files |
5 | Run the same assignment again after a month and compare | Retains context across sessions, which is what makes the comparison possible |
Sample prompt:
Context: I am running a small room with two named agents and myself. I need delegation that can be reviewed member by member rather than as one answer. Role: AI as one named team member with a written brief. I own every decision and the record. Profile: Act as a Co-Worker and Assistant, the one this task needs. Take the assignment, execute it, and report back to the room. Task: As the researcher, produce (1) a source list with one line on why each source qualifies, (2) the five findings that change the decision in front of us, (3) the finding you are least confident in, and (4) your recommended next step for the analyst. Constraints: Stay inside the boundary I wrote for this role. Do not speak for the other member. Do not produce the analyst's output. Name what you could not establish. Output format: A short brief addressed to the room, ending with a section headed Open questions for the analyst. Memory: Decisions and boundaries are recorded in my own project file, because the room's memory is not a record of accountability. UP-Context verification: I reviewed each member's output against its own written boundary before I used any of it. Data safety: this pack carries no personal data. I do not put personnel matters, client identities or unpublished financials into a room that other members and the vendor's infrastructure can see.Verification checklist:
☐ Multi-Model Check: for a high-stakes assignment, have one member's output reviewed by a member you did not use for it.
☐ External Source: check any factual claim that will leave the room against a source outside it.
☐ Human Review: a colleague reads the room's output for whether the attribution is clear, not for whether the work is good.
☐ CI-First Test: can you state which member produced which part, and defend each part on its own? If not, the room is producing an unattributed average.
The verification rule that covers all three
Every workflow above ends the same way, and the order is not optional. Run the multi-model comparison on the material that matters. Open the external source. Have a human read anything that leaves your hands. Then apply the CI-First test, which is the only one that measures you rather than the tool: if you cannot explain and defend the output without the assistant in front of you, the output is not ready, however finished it looks.
Strengths, Limits, and AI Imposture Risk
Strengths
The tool delivers clear CI-First benefits in these areas:
CI-First Benefit | Strength | Evidence |
Time | Real savings on first attempts, and unattended execution is the mechanism rather than a claim | A described task runs to a finished artefact without supervision, scheduled tasks and background check-ins run on a clock, and the vendor documents parallel task execution inside one project. The saving is structural: the work happens while you do something else, which no chat assistant provides |
Quantity | Strong, and it is the clearest advantage | One account produces research pages, decks, sheets, documents, dashboards, images, video, audio and drafted mail from the same instruction style. The vendor's own catalogue lists more than 110 named generators, and the team surface multiplies by the number of named agents a person runs |
Quality | Marginal to moderate, and the score is set by the absence of measurement rather than by observed weakness | The vendor publishes no accuracy, hallucination or task-completion figure for the agent surfaces, no independent evaluation exists, and the one benchmark figure in circulation is a self-report from 2025 that is absent from the official leaderboard. Output is formatted to a professional standard, which raises the cost of finding the errors rather than lowering them |
Skill | Marginal | The product teaches by producing, and the memory layer removes the need to hold your own context. Where a reader learns, they learn verification discipline, which is real and valuable and is the same lesson every agentic tool in this series teaches. Nothing in the product builds the underlying craft it performs |
Limits
The tool is weak or brittle in these areas:
No published measure of correctness. No task-completion rate, no hallucination rate, no accuracy figure for the agent surfaces, and no independent evaluation of them. Everything about verification rests on the user, and the output style makes verification harder rather than easier.
A meter that prices after the fact. The headline capability carries no published per-run cost, credits do not roll over, an upgrade that changes the billing interval revokes unused credits, and failed runs can consume the balance. The billing complaint record in the public reviews is dense and specific.
A memory layer with no write-approval step and no routine read-back. Context is compiled from connected sources automatically, and there is no per-write approval and no routine practice of reading what was compiled. That is the mechanism behind the Skill Illusion rating of High.
Session limits that sit behind promotional language. Users report a recurring five-hour session window, the vendor acknowledged a session-limit change publicly in April 2026, and the promotional arrangements for chat and image generation carry stated end dates.
Two pricing surfaces that disagree. The business pricing card and the help centre print different credit allocations for the same Team and Enterprise plans, both read on the same day.
A pricing page that requires an account. The rate card is not readable without signing in, so published figures from third parties are approximations by definition.
A cloud computer that is erased on cancellation. The vendor states that data on it cannot be recovered, so the persistent assistant's state is not durable unless you keep your own copy.
Prompt injection on the surfaces that act for you. The vendor publishes a warning for the persistent assistant that content it reads may attempt to redirect it, which is an honest disclosure of a live risk rather than a solved problem.
AI Imposture Risk
Time Illusion
Rating. Medium
Evidence. The savings are real where the task is well suited: unattended execution across a decomposed task is genuine net gain rather than a claim. The trap sits in the verification tax and the meter. The output arrives formatted to a professional standard with no reliability figure attached, so checking it is slow, and every check consumes the reader's time after the tool has already consumed credits. Users report the meter and the session window as the two things that turn a fast task slow, and the vendor publishes no per-run price to plan against
Quantity Illusion
Rating. Medium
Evidence. The volume is unusually high and the interface makes shipping unverified output easy: decks, briefings and drafted replies arrive finished, and the surfaces that draft in your own voice are the ones most likely to be sent without a full read. The mitigating evidence is structural rather than cosmetic: research pages carry citations, so a reader has something to check, and the product documents where it could not reach rather than filling the gap silently. The trap is not that the volume is bad; it is that the volume is plausible enough to skip the check
Skill Illusion
Rating. High
Evidence. Two mechanisms, both documented. First, the workspace performs the craft at every institute at once: it researches, writes, designs, analyses and drafts communication, so a user can produce work in domains where they hold no skill and cannot evaluate the result. Second, the memory layer compiles the user's context automatically from connected sources. Applying the framework's clause 5.2.3-a as written: a human approves no individual memory write, and there is no routine practice of reading what was compiled, which are the clause's own two escalation conditions. The floor is satisfied and exceeded
Overall Imposture Risk: Medium
Framework 5.3 gives Medium where one trap is High with clear mitigations, or where one or two traps are Medium. Here one trap is High and two are Medium, and the mitigations are real: sources are cited on the research surface, the product documents its own unreachable areas, the persistent assistant warns about prompt injection, and local execution is offered for the surface that touches the file system. The discipline the framework requires is specified in the Superhuman Usage Guidance below.
Framework v1.2 clause note
Clause 5.2.3-a, agent-authored procedural memory, with a Skill Illusion floor of no lower than Medium, applies. The tool writes the user's memory on the user's behalf: the memory layer compiles context from mail, calendar, chat, documents and the workspace's own project history, and the persistent assistant holds long-term memory across sessions and channels that the vendor describes as learning the user's preferences. No per-write approval step exists on that surface, and no routine practice of reading what was compiled is documented; the help centre describes connection and disconnection of sources, not review of what was written. Both escalation conditions in the clause are therefore met, the floor is exceeded, and the trap is rated High.
Clause 4.2-a, agent-mediated conversation, applies. The product drafts mail in the user's voice, and the persistent assistant can send messages and act through messaging channels under the user's own accounts, which is condition (a) in the clause: agent-authored text presented as the person's own voice in a human-facing channel. The clause's neutral case does not obtain here, because the agent is not disclosed as itself in the channels where it acts on the user's behalf. The mitigation is stated in the Superhuman Usage Guidance: read anything consequential before it leaves.
Clause 7.5, team-level rooms, applies. The collaboration surface is a shared room in which named agents and the human work together, and the vendor documents agents holding roles, filing tasks and delivering work into shared group chats. Under the framework a room with more than one agent requires Centaur as the default, with a written task boundary per member and review of each member's output, and Cyborg is not available for such a room. Centaur is assigned in this review, and Workflow 3 above is built to the clause's requirement.
No framework release is requested. All three clauses are applications of clause text that already exists in framework v1.2.
Section 7c: The data position, the training terms, and the money clauses, stated plainly
This section reads the vendor's own documents against each other. It changes no sub-score, and it says why at the end.
Who you contract with, and how the documents relate. The terms of service name MainFunc Inc. together with its subsidiaries and affiliated entities, including Genspark Inc. as a wholly owned subsidiary. The terms were last updated on 2 April 2026 and apply to individual, team and enterprise customers alike, except where an enterprise contract supersedes them on a conflict. Two clauses in the opening section are worth reading before the rest: the vendor reserves the right to restrict, suspend or terminate access at any time and for any reason, and continued use after a posted change constitutes acceptance of the changed terms.
What is collected. The privacy policy, last updated 24 July 2026, states that the service collects automatically, among other things, the prompts you input and the resulting outputs, and that tracking technologies are used to analyse and improve the service. It states that queries may be transmitted to third-party model providers to serve requests, names OpenAI and Anthropic as examples, and includes the Google Workspace limited-use affirmation that Workspace data is not used to develop, improve or train generalised models. Read the collection list carefully: the prompts and the outputs are named as collected data in their own right, rather than as transient processing.
The training position, and the gap between tiers. The business pages state that customer content is never used to train models, "contractually and technically", and the team and enterprise help page confirms that both paid team tiers are automatically opted out and that subprocessors are bound by no-training agreements with zero data retention where applicable. The individual plan surfaces, including the consumer terms and the privacy policy, do not state the same position in the same terms. That is a real asymmetry and it is the single most consequential governance point in this review for a U365 reader on a personal plan. The safe reading, given the documents, is that the strongest published commitment applies to the business tiers, and a reader on Plus or Pro should treat the individual position as not stated rather than as equivalent.
The memory layer, read as a data-processing decision. The memory layer compiles mail, calendar, chat, documents, meeting notes and project history into a standing context that every subsequent task reads. The vendor states that the compiled data is accessible only to the account holder, that integrations require explicit authorisation and can be disconnected at any time, and that the product holds enterprise certifications. What the documentation does not state is a per-write approval on the compiled context, an export of what was compiled, or a retention rule for the context itself. The practical consequence is that the most useful surface is also the one whose contents the user has no routine practice of reading, which is the same finding the clause note records from the pedagogical side.
The persistent assistant and the cloud computer. If you take the cloud computer rather than local execution, note the two positions the vendor states plainly. The cloud computer is a dedicated machine with a fixed address, and the vendor states that on cancellation it is reclaimed and all data on it is permanently erased and cannot be recovered, so critical files must be copied out first. The desktop application's local mode is documented with a security note that a chosen workspace folder is soft guidance rather than a hard boundary, and that backing up files before any file-related task is advisable. A vendor that writes those two sentences has told the reader where the sharp edges are.
The recording surface. The memory layer's hardware companion records conversations in a room. The help centre publishes a consent guide that states recording law varies by country and by state, that some places require consent from all participants, that people should be notified before recording starts and given a real chance to decline, and it lists the conversation classes to avoid: healthcare, legal and financial discussions, human-resources matters, student situations, and any place where people reasonably expect privacy, closing with "When in doubt, do not record." That guide is a genuinely good document, and it is the vendor's own advice that the reader must apply.
The money clauses. Read together, they are specific.
Credits do not roll over. Unused credit at the end of a cycle is not carried forward, and a downgrade resets the allocation rather than preserving it.
An upgrade that changes the billing interval revokes unused credits from the previous plan and issues a fresh set under the new one.
Refunds: one refund per account, ever, and a second request cannot be processed after the first is approved. For most customers the window is 24 hours after purchase on a monthly plan and 72 hours on an annual one. Customers in the European Union, the United Kingdom and Turkey have 14 days, and customers in Korea have 7 days, on both monthly and annual plans.
The vendor may reject a refund where the system detects excessive use of the service or of a quota before the request, and it states that position in the same section as the window.
Commercial-use rights are stated for paid members, and the two surfaces disagree on the date: the help centre guarantees them through 30 June 2027 for Plus and Pro, while the team plan page states 31 December 2026.
Promotional arrangements carry stated end dates. Core chat and image generation are included at no credit cost and the arrangement for pre-existing subscribers runs unchanged through 31 December 2026, after which the plan terms then in effect apply.
Why no sub-score moved. Section 7c records governance findings, and the framework does not grade a vendor's contract drafting. The training asymmetry between tiers, the credit rules and the refund windows are facts a reader needs in order to decide whether to buy; they do not change Time, Quantity, Quality or Skill, and the Humics ratings and the trap ratings are set by the mechanisms described under Strengths, Limits, and AI Imposture Risk above. Two of these findings do feed the review: the credit rules are part of why the Time trap is rated Medium, and the memory position is part of why the Skill trap is rated High. Neither is moved by this section; both are already counted where they belong.
What this section does not do. It does not constitute legal advice, and it does not resolve the tier asymmetry; it records it. A reader whose institution requires a no-training commitment, a residency guarantee or a data-processing agreement should treat the enterprise tier as the starting point for that conversation, because that is where the vendor publishes those commitments, and should read the terms as they stand on the day rather than as this review describes them on 2026-09-28.
U365 Co-Intelligence Rating
CI-First Profile
Primary profile: Co-Worker and Assistant (2). The product's core promise is execution: it takes a described task and returns finished work, which the user then reviews and uses.
Secondary profiles: Analyst and Tester (4), on the research and fact-check surfaces, which gather, compare and report rather than produce. Co-Creator and Thought Partner (1), narrowly, on the design and build suites, where a described concept becomes an artefact that can be iterated.
The other two profiles are not assigned. The product is not a Coach or Tutor (3): it produces rather than teaches, and the Skill rating records that. It is not a Challenger or Devil's Advocate (5): it answers the brief it was given, and a user who wants to be contradicted has to build that instruction deliberately, which the workflows above do.
CI-First Benefit Score
Time
Score (0-10). 6
Rationale. Moderate net savings where the task fits, and the mechanism is structural rather than claimed: unattended and scheduled execution is work that happens while you do something else, which no chat tool provides. Held at 6 rather than higher because the verification tax is real and the meter is unreadable in advance. A tool that produces finished-looking artefacts with no published reliability figure makes checking slow, and the session window plus the credit draw can turn a fast task slow in the same session. The saving is moderate and consistent on first attempts, which is the scale's own description of 6
Quantity
Score (0-10). 7
Rationale. Strong increase in usable output per hour of human work, and it is the product's clearest advantage. One account covers research, decks, sheets, documents, dashboards, images, video, audio, drafted mail and internal tools, and the team surface multiplies the count by the number of named agents running. Scored 7 rather than 8 because the usable half of the output depends on the reader's own verification, and the credit balance bounds the volume on every plan
Quality
Score (0-10). 4
Rationale. Marginal, and it is scored from measured absence rather than from measured weakness. The vendor publishes no accuracy, hallucination or task-completion figure for the agent surfaces; no independent measurement exists; the one benchmark figure in circulation is a self-reported 2025 claim that does not appear on the official leaderboard for that benchmark. Output is formatted to a professional standard, and polished formatting raises rather than lowers the cost of finding the errors. 4 is the score the framework produces when the category publishes no measurement at all
Skill
Score (0-10). 4
Rationale. Marginal, and scored conservatively for the reasons the framework gives. The product performs the craft rather than teaching it, and the memory layer removes the need to hold your own context, which is the mechanism behind the High Skill Illusion rating. It is not lower because two genuine skill paths exist: verification discipline, which any sustained user of this class of tool develops, and the delegation discipline the room surface forces a user to learn by making the agents and their tasks visible
CI-First Benefit Score: 4.8 / 10 (CI-First Positive)
Why this score is not higher, and why it is not lower
4.8 is CI-First Positive. The band label matters and it should be read with the number: this is a recommendation for the work it fits, not a warning. A person producing internal briefings, decks, first drafts and asset batches on a recurring basis has a genuine net benefit available here, and two of the four dimensions are at or above 6.
The score is not higher for two reasons that compound. Nothing independent measures whether the agent's work is correct, and the cost of a run is disclosed after it ran rather than before. A tool that returns finished artefacts you cannot always judge, on a meter you cannot price in advance, cannot reach the Strong band on the framework's rules, however impressive the demonstration is, and the demonstration is impressive. Quality at 4 and Skill at 4 are the two dimensions holding the total down, and both are the dimensions the framework scores most conservatively by design: the Quality score records an absence of measurement across the whole agent category rather than a defect of this product alone, and the Skill score records that a workspace which performs the craft cannot also be the thing that teaches it.
The score is not lower because the core capability is real, broad and structurally delivered. Unattended execution, parallel task handling, a working memory layer, an office suite that produces real files, and a team surface where the delegation boundary is visible are all beyond what most tools in this series deliver. The vendor also documents its own sharp edges: the prompt-injection warning, the erasure on cancellation, the local-mode boundary note, and the recording-consent guide. A vendor that publishes those is not hiding the risk it asks a reader to accept.
Humics Protection Badge
Dimension | Rating | Rationale |
Creativity | Neutral (0) | The product generates ideas and artefacts in quantity, and the design and build suites turn a concept into something concrete. It does not replace the user's own ideation in the common case, because the brief is still the user's, and it exposes a user to a wide range of styles and forms they can compare. It does not strengthen creativity either: nothing in the product trains a user to originate, and the volume of available generation makes it easy to select rather than to invent |
Critical Thinking | Erodes (-1) | Three mechanisms, all documented. The output arrives finished and formatted, which is the presentation most likely to be accepted without inspection. The product publishes no reliability measure, so the user has nothing to calibrate trust against, and the vendor's own research surface cites sources in a way that reads as verified. And the persistent assistant warns that content it reads may attempt to redirect it, which places a judgement the user cannot easily make, distinguishing instructions from content, inside an automated reading loop |
Social Authenticity | Erodes (-1) | The product drafts communication in the user's own voice from the user's own mail, and the persistent assistant can send on the user's behalf through accounts that recipients read as the person themselves. The vendor states the assistant should be transparent about being an assistant, and the mailbox help text states that the reader's voice and history power the drafting, which makes the authorship question sharper rather than softer: the better it performs at sounding like you, the less the recipient can tell. The mitigation is disclosure and human reading of anything consequential, which is the reader's discipline rather than a product feature |
Humics Protection Score: -2 / +3 Badge: Humics-Risky
This is the twelfth Humics-Risky record in the series, measured against the live Tools collection on 2026-09-28 rather than carried from the register. It is not awarded for tone. It is awarded because two of the three dimensions are eroded with no compensating protection, and because the mechanism in the Critical Thinking column, a product that returns finished work with no published measure of its correctness, is the same class of finding the framework exists to record.
Superhuman Usage Guidance
When to invite this tool:
Research synthesis on a topic you already understand well enough to check the sources.
First drafts of internal artefacts: decks, briefings, summaries, spreadsheets, campaign material.
Recurring operational work with a clear finish line: scheduled reports, monitoring, meeting notes and follow-ups.
Media production at volume where the output is reviewed before use.
Team delegation where the agents hold written boundaries and each member's output can be reviewed on its own.
When to keep this tool out:
Any deliverable you cannot verify yourself, especially in a domain where you hold no skill.
Anything sent in your voice to a human being without being read, in full, by you.
Material under a confidentiality obligation, unless the tier you are on carries the commitments your institution requires.
Decisions with legal, medical, financial or safety consequences. Nothing in this product publishes a reliability measure that would justify that use.
Work whose cost has to be known before it starts.
U365 method integration:
LIPS + CARE: The memory layer is the closest commercial analogue to LIPS in this series, and the discipline transfers directly: what enters the record is a decision, not a default. The product is strong at Collect and Execute and silent at Review, so the Review step is exactly where the reader must place the effort.
ULM + EVA: The product supports the Career and Quality of Life domains most directly, through the artefact work and the automation of recurring administration. It has no role in the Body, Spirit or Social domains, and nothing in it exercises Character, so the 6-domain model stays the reader's own instrument.
UP-Context: The product responds well to structured instructions, and the workflows above use the house prompting method deliberately: role, profile, task, constraints, and the two closes that matter most with this tool, the memory disclosure and the data-safety boundary.
SL-OS: Outputs are files, which means they can be filed in the reader's own system rather than left in the workspace. That is the integration path: the product generates, the reader's record keeps.
UNOP: The product has no pedagogical mechanism by design. Applied to learning, it should be used to produce practice material for the reader's own review cycle, never as a substitute for the retrieval and reflection steps.
Over-delegation warning: Over-delegation with this tool has a specific shape, and it is different from the usual warning. The usual failure is that you ship unverified work. The failure here is that you stop holding your own context. The memory layer compiles what you know, the agent executes what you would have done, and the artefacts come back formatted well enough to pass. Six months later you have a body of work you cannot explain, in domains you cannot work in without the subscription, and a record of your own thinking that you never read. The framework's formula is the test: if your Human Intelligence drops while the tool keeps producing, the Co-Intelligence outcome falls with it, however good the output looks. The practical rule is the one the workflows above encode: read what the memory compiled before you add a source, verify one item per week by hand, and keep a copy of what matters outside the account.
What Users Say
Aggregate Rating Table
Platform | Rating | Number of reviews | Link |
Google Play, Genspark AI Workspace | 4.5 out of 5 | About 101,000 | |
Apple App Store, Genspark AI Workspace (United States) | 4.71 out of 5 | 4,244 | |
Trustpilot | 3.8 out of 5 | 443, of which 434 in the last 12 months | |
G2 | 3.7 to 3.8 out of 5 | Five to seven, depending on the reading | No single stable public aggregate; read qualitatively only |
Product Hunt, browser launch | 4.0 out of 5 | 11 | |
Product Hunt, Word plug-in launch | 4.45 out of 5 | 11 | |
Reddit, community sentiment | Mixed | Multiple recurring threads |
Capterra, GetApp, Futurepedia and FutureTools carried no aggregate for this product on 2026-09-28.
What Users Praise
Three themes repeat across every surface that carries volume. First, breadth: reviewers describe the workspace as a one-stop shop and say the range of tools in one place removed other subscriptions. Second, speed on research and presentation work: the research pages and the slide generator are the two features named most often, and reviewers describe hours of research or deck-building collapsing into minutes, including one reviewer using the design surface for architectural renders and another building two weeks of social content from one session. Third, output quality relative to effort: Trustpilot's own summary of 304 recent reviews describes most reviewers as "somewhat happy", with convenience, time savings and the breadth of tools as the dominant positives, and the platform's own reply behaviour (it replies to some reviews) is visible on the profile.
What Users Complain About
One theme dominates, and it is not a matter of taste: the credit system. Reviewers report credits draining faster than expected on heavy tasks, credits consumed by failed runs, and difficulty tracking which task consumed what. A Google Play reviewer describes four attempts at one video generation ending in backend errors and an exhausted balance. A Trustpilot reviewer paying 28 dollars a month reports eight days of a broken image-generation panel where the reset timer advanced by five hours on every attempt, and describes support replies as templated. Others report annual subscriptions continuing after they no longer had a working service, refund requests going unanswered, and support response times measured in weeks. The historical record sharpens the point: earlier in 2026 the Trustpilot profile sat near 1.5 to 1.7 out of 5 with a very high share of one-star reviews, and community threads from that period describe users hitting session limits that were not in the terms when they subscribed. The current 3.8 is a real recovery, and it is a recovery from that position rather than a long record of satisfaction.
A second, smaller theme is reliability of the more ambitious surfaces: reviewers report generated output that needed substantial manual rework, image quality that changed between releases, and features advertised as unlimited hitting a session limit that resets on a fixed window.
Sentiment Summary
Overall sentiment: Mixed, and deliberately described that way rather than as either of its halves. The volume of positive sentiment is large and on the biggest surfaces. The negative sentiment is concentrated on one thing, it is loud, and it is about money rather than about capability.
Key themes:
Breadth and time savings on research and presentation work are praised consistently and across platforms.
Credit consumption, session limits and refund handling are the dominant complaints, and they recur on every surface.
The gap between the mobile-store ratings and the paying-customer review sites is wide, and it reflects two different populations rather than a disagreement about the same experience.
Support responsiveness is the second-most common complaint, with templated replies and long waits described repeatedly.
Failed runs consuming credits is a specific, repeated pattern rather than an isolated report.
U365 Editorial Note
The crowd sentiment and the CI-First evaluation agree on the mechanism and disagree on its importance, and the disagreement is worth naming. Users are not complaining that the tool does not work; they are complaining about the meter, which is exactly where this review found the Time Illusion to be Medium rather than Low: the capability is real and the cost of discovering it is unpredictable. The reviews are the field evidence for a scoring decision that would otherwise read as theoretical.
The same applies to the Skill Illusion rating of High. No reviewer complains about the memory layer, and none of them would: a feature that removes the need to brief the machine every session feels like an improvement, and the cost, which is that you stop holding your own context and never read what was compiled, is invisible from inside the product. That is what an illusion is, and it is why the framework rates it rather than the market. The one place the users and the framework converge without any interpretation at all is the output's credibility: reviewers praise how finished the work looks, and the framework's whole finding is that nothing published says whether it is right.
Comparison and Alternatives
This product's comparison set is unusually wide, because the workspace competes with the bundle a person already has rather than with one tool.
Alternative | Choose that instead if you need... | Choose Genspark if you need... |
A single frontier assistant on a monthly subscription | One model you learn deeply, a predictable flat price, and work that stays in a conversation you are reading | Work delivered as files, several capabilities under one balance, and execution that runs without you |
A dedicated research tool | Verifiable, citation-first answers as the whole product, with no generation surfaces to distract from it | Research as one step inside a larger deliverable: the briefing becomes the deck becomes the campaign |
A dedicated deck or design tool | Craft-level control over a small number of artefacts | Volume across many artefact types, with the first pass produced for you |
A self-hosted agent stack | Data residency, model choice, and full inspection of what runs on your hardware | Setup measured in minutes, vendor-maintained integrations, and no operational burden |
An office-suite add-in from the platform vendor | The assistant inside Microsoft 365 or Google Workspace with the vendor's own compliance posture and no third party in the chain | A wider capability set than the plug-in offers, at the cost of a third-party processor holding the context |
Where Genspark is clearly better: breadth with execution. Very few products in this series take a described task and return a finished, editable artefact across research, documents, spreadsheets, presentations, media and code, and fewer still run it unattended on a schedule. The memory layer compounds this: once sources are connected, every later task starts with context the user did not have to supply, which is the difference between a tool you operate and a workspace you inhabit.
Where Genspark is clearly worse: cost predictability and verifiability. A single-purpose assistant on a flat subscription tells you what it costs and keeps one conversation in front of you. This product prices its headline capability after the run, does not roll credits over, and hands you artefacts whose correctness nothing published measures. On those two axes, the narrower tools are the safer choice, and this review does not argue otherwise.
Verdict and Next Steps
Who should adopt it: Anyone whose work is internal, reviewable and repeatable: researchers assembling briefings, marketers producing content at volume, operations teams automating recurring reports, and small teams who want agents with visible roles. Not for anyone whose work cannot be checked, sent, or costed before it happens.
When: After you have run one real task on the free plan and read its sources back. Do that before you connect an inbox.
For what: Turning instructions into finished artefacts, and running the same instruction again next week without rewriting it.
UP-Context prompt pack:
The verified briefing Context: I am preparing material I will have to defend in writing and in conversation, in a subject where the readers know more than I do. Role: AI as a research assistant. I own the judgement about what is true and the final wording. Profile: Act as an Analyst and Tester, the one this task needs. Test the material rather than asserting a position. Task: Produce (1) the question restated in one sentence, (2) the three positions with real support and the evidence for each, (3) the evidence that weakens each, (4) where sources disagree and why, and (5) what would change my mind. Constraints: Cite a source for every factual claim, name what you could not source instead of asserting it, and do not average conflicting sources into one position. Output format: A headed outline of at most 1,200 words, ending with a section headed What I could not establish. Memory: The verified version goes into my own record, because the tool holds no standing instruction of its own. UP-Context verification: I opened every cited source and removed every claim I could not trace to one. Data safety: this pack carries no personal data. I do not paste unpublished institutional material, client names, or anything under a confidentiality obligation into it.
The delegation brief, one agent at a time Context: I am running a shared room with named agents and myself, and I need to review each member's work on its own. Role: AI as one named team member with a written brief. I own the decisions and the record. Profile: Act as a Co-Worker and Assistant, the one this task needs. Take the assignment and report back. Task: Produce (1) the sources you used and why each qualifies, (2) the findings that change the decision, (3) the finding you trust least, and (4) what the next member should do. Constraints: Stay inside the boundary I wrote for this role, do not speak for the other members, and name what you could not establish. Output format: A short brief addressed to the room, ending with a section headed Open questions for the next member. Memory: Boundaries and decisions are recorded in my own project file, because the room is not a record of accountability. UP-Context verification: I reviewed each member's output against its own written boundary before I used any of it. Data safety: this pack carries no personal data. I do not put personnel matters, client identities or unpublished financials into a room that other members and a vendor's infrastructure can see.
The cost-and-scope check before a run Context: I am about to spend credits on a task whose price I cannot see in advance, on a balance that does not roll over. Role: AI as a planning assistant. I own the decision to run it and the budget it has to fit. Profile: Act as an Analyst and Tester, the one this task needs. Estimate and bound rather than reassure. Task: State (1) what this task will actually produce, (2) which parts are mechanical and which need judgement from me, (3) the cheapest version of it that would still be useful, (4) what would make you stop rather than continue, and (5) what I should check first to avoid paying for a run I did not need. Constraints: Do not estimate a credit figure you cannot support. Say what is unknown instead. Output format: A short plan of at most 400 words, ending with a section headed What I cannot price. Memory: I record the actual consumption after the run in my own cost log, because the tool keeps no standing budget of mine. UP-Context verification: I checked the balance before and after the run, and I wrote down what it cost. Data safety: this pack carries no personal data. I do not paste anything into it that I would not put in a third-party account.
Related U365 content:
The INSIDE Tools library on university-365.com carries the full series of CI-First reviews, including the other agent and workspace platforms reviewed this month.
The published Online Programs catalogue carries the programmes named in the Tool to Skill to Credential chain above, each with its own duration and step count.
Status and Last Tested
Status: Active Last tested: 2026-09-28 Version tested: Genspark AI Workspace 6.0, as documented at genspark.ai, in its help centre, terms, privacy policy and business pages in September 2026. The product ships continuously and publishes no version number on its own surfaces, so the documentation date is the tested reference. Next re-test: trigger-based, maximum six months. Re-check triggers: the eight listed in the Status and Re-check block at the top of this review, of which the two most consequential are a published quality measurement of the agent surfaces and any change to the credit model or the promotional pricing arrangement.
What Active means here. Active is the status for a tool that is current and recommended, and this review uses it deliberately. The product is in active development, widely adopted, and delivers a real benefit for the work it fits; the recommendation in the Verdict section is a real one. Active does not mean unqualified: the Imposture Risk is Medium with a High Skill Illusion, the Humics badge is Risky, and the conditions that make the recommendation safe are listed in the same section.
Migration Path
Not applicable. Genspark is Active, and no migration is recommended or required.
This section is retained so the review's structure matches the rest of the series, and it states plainly that a reader who adopts Genspark has no migration obligation. A reader who wants to leave should note two things before cancelling: the cloud computer that hosts the persistent assistant is erased once the cancellation takes effect, so files must be copied out first, and credits remaining on the account do not carry to any other plan or product.
U365's Recommendations to Learn More
Official learning resources
Genspark help centre: https://www.genspark.ai/helpcenter
Super Agent documentation: https://www.genspark.ai/helpcenter/super-agent
Membership plans and credit rules: https://www.genspark.ai/helpcenter/membership-plans
SecondBrain documentation: https://www.genspark.ai/helpcenter/secondbrain
Persistent assistant documentation: https://www.genspark.ai/helpcenter/genspark-claw
Team and enterprise terms: https://www.genspark.ai/helpcenter/team-enterprise-plans
Vendor blog, including the Workspace 6.0 announcement: https://www.genspark.ai/blog
Video tutorials and channels
The vendor's own channel publishes short product walkthroughs, including one on the memory layer, at https://www.youtube.com/@gensparkai
Two vendor walkthroughs that were checked and resolve on 2026-09-28: the memory layer introduction at https://www.youtube.com/watch?v=Ty4ybR233l4 and the design surface introduction at https://www.youtube.com/watch?v=Bal6aLWLBRY
Third-party tutorial channels cover the workspace surface comprehensively; treat their claims as commercial content and check any performance claim against the product itself.
Two vendor walkthroughs are embedded below, both checked and resolving on 2026-09-28. The first is the memory layer, which is the surface this review rates as the highest risk; the second is the design surface, which is the clearest short demonstration of what the workspace produces. Neither substitutes for running one real task on the free plan.
Written tutorials and deep-dive articles
Pricing analysis, with the credit costs per task class: https://www.eesel.ai/blog/genspark-ai-pricing
An independent three-week hands-on review covering the workspace surface: https://www.lindy.ai/blog/genspark-review
Coverage of the vendor's funding, growth and the Microsoft partnership: a June 2026 wire report at https://www.channelnewsasia.com/business/ai-startup-genspark-valued-26-billion-in-latest-funding-round-6190486
Community and social
Reddit community, which is the most candid surface on billing: https://www.reddit.com/r/genspark_ai/
Google Play listing and its review stream: https://play.google.com/store/apps/details?id=ai.mainfunc.genspark
Trustpilot profile: https://www.trustpilot.com/review/genspark.ai
Resources on Genspark
The vendor's own account is the first channel to add, because a change to the credit rules, the pricing ladder or the training position would be announced there before it reached a documentation page.
Resources on X
Dedicated X channels: the vendor's own account, @genspark_ai, carries product announcements and release notes. Independent accounts discuss the product's pricing changes more critically, and the community subreddit below is the candid surface on the same subject.
CI-First Evaluation Summary Card
Genspark (Genspark AI Workspace) Evaluated: 2026-09-28 Framework: U365 CI-First Evaluation Framework v1.2
CI-First Benefit Score | 4.8 / 10 (CI-First Positive) |
Time | 6 |
Quantity | 7 |
Quality | 4 |
Skill | 4 |
Humics Protection | Humics-Risky (-2 / +3) |
AI Imposture Risk | Medium (Time Illusion Medium, Quantity Illusion Medium, Skill Illusion High) |
CI-First Profile | Primary Co-Worker and Assistant (2); secondary Analyst and Tester (4) and Co-Creator and Thought Partner (1) |
Collaboration Mode | Centaur |
Status | Active |
Last tested | 2026-09-28 |
In one sentence: A broad, genuinely capable workspace that turns instructions into finished work across research, documents, media and team operations, sold on a meter that prices after the run and with no published measure of whether the run was right.
Glossary
CI-First
Co-Intelligence First. The U365 principle that human intelligence is the ruler and the orchestrator, and artificial intelligence is the amplifier. A tool is judged by whether the combination is more profitable than human intelligence alone, not by what the tool can do.
CI-First Benefit Score
The arithmetic mean of the four benefit dimensions, Time, Quantity, Quality and Skill, each scored from 0 to 10. Bands: 0 to 2.0 CI-First Negative, 2.1 to 4.0 CI-First Neutral, 4.1 to 6.0 CI-First Positive, 6.1 to 8.0 CI-First Strong, 8.1 to 10.0 CI-First Transformative.
Time Benefit
How much time the tool saves net of the time spent instructing it, reading its output, verifying it and correcting it. A tool that saves thirty minutes of work but costs twenty-five minutes of checking and fixing scores on the net five.
Quantity Benefit
How much more usable output a person produces in the same time, counting only output that survives verification. Volume that cannot be checked is not counted here; it is counted in the Quantity Illusion.
Quality Benefit
Whether the output is better than the person would have produced alone, measured after verification rather than by appearance. A tool that produces polished output containing subtle errors has not delivered a quality benefit.
Knowledge and Skill Benefit
Whether the user learns something durable, or whether the tool performs a skill the user does not hold and does not acquire. This is the dimension the framework scores most conservatively, because it is the one where an illusion is easiest to sustain.
CI-First Profile
Which of five roles the AI plays in the relationship: Co-Creator and Thought Partner, Co-Worker and Assistant, Coach and Tutor, Analyst and Tester, or Challenger and Devil's Advocate. The profile is attributed before a task is given, not after.
Humics
The three human capabilities the framework tracks, from Pascal Bornet's work: Creativity, Critical Thinking and Social Authenticity. The rating asks whether sustained use of the tool strengthens the human or atrophies them.
Humics Protection Badge
The sum of the three Humics dimensions, each rated +1 protects, 0 neutral, or -1 erodes. 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 user in one of three illusions: the Time Illusion, where the appearance of speed hides the cost of prompting and checking; the Quantity Illusion, where the volume of plausible output hides its quality; and the Skill Illusion, where the appearance of competence hides that the user is not learning. Overall risk is Low when all three traps are Low, Medium when one or two are Medium or one is High with clear mitigations, and High when two or more are High or the core promise rests on an illusion.
Centaur
The collaboration mode in which the split between human and machine is defined before the work starts and each side's output is reviewed by the other. It is the framework's default because it is the safest mode for most tools and most users.
Agent-authored procedural memory
Memory, skills or standing instructions written by the agent on the user's behalf rather than by the user. It is treated as a Skill Illusion vector in its own right, because the user ends up holding a durable capability they did not write. The framework sets a floor of no lower than Medium on the Skill Illusion for any tool that writes memory for the user, and rates it High where the agent can revise that memory during use without a per-write human decision, or where the user has no routine practice of reading what was written.
User Sentiment
The aggregated public opinion from the review platforms, community forums and directories. It is reported separately from the CI-First score because crowd sentiment can contradict a rigorous evaluation: where the two agree the finding is stronger, and where they diverge the divergence is worth explaining. For Genspark the four surfaces are read as four different instruments rather than averaged: Google Play at 4.5 from about 101,000 reviews, the App Store at 4.71 from 4,244 ratings, Trustpilot at 3.8 from 443 reviews, and G2 at 3.7 to 3.8 from five to seven reviews, with Capterra, GetApp, Futurepedia and FutureTools carrying no aggregate at all.
Review Status
Review Status records the current standing of the tool at the time of the last test. Active: the tool is current and recommended. Active (updated): recently re-checked and the content was refreshed. Changed: a re-check trigger fired and an update is pending, so read the review with that in mind. Risky: the tool has significant unresolved issues, or it has been clearly surpassed by newer alternatives. Use it with caution and read the Limits section. Retired: the tool still works but is no longer recommended. Deprecated: the tool has been shut down or fundamentally changed. Retired and Deprecated posts include a Migration Path section.
Last tested and Re-check
Last tested is the date on which the vendor's own surfaces, as they stood on that date, were read to establish every fact in this review. Re-check is trigger-based with a six-month ceiling: the review states what would change the assessment, and the tool is re-read when one of those triggers fires or when six months have passed.
Sources
Vendor primary sources, all read 2026-09-28
Genspark home page and product surfaces: https://www.genspark.ai
Workspace 6.0 announcement: https://www.genspark.ai/blog/genspark-ai-workspace-6
Terms of service, last updated 2 April 2026: https://www.genspark.ai/terms
Privacy policy, last updated 24 July 2026: https://www.genspark.ai/privacy
Help centre, membership plans, credits and subscription management: https://www.genspark.ai/helpcenter/membership-plans
Help centre, Super Agent: https://www.genspark.ai/helpcenter/super-agent
Help centre, SecondBrain: https://www.genspark.ai/helpcenter/secondbrain
Help centre, SecondBrain Note and its recording consent guide: https://www.genspark.ai/helpcenter/secondbrain-note
Help centre, persistent assistant including the prompt-injection warning and the local-mode note: https://www.genspark.ai/helpcenter/genspark-claw
Help centre, Team and Enterprise plans, including the training position: https://www.genspark.ai/helpcenter/team-enterprise-plans
Business pages, including the security and governance claims: https://www.genspark.ai/business
Trust page: https://trust.genspark.ai
Device store: https://shop.genspark.ai
Independent sources
Agency wire report on the June 2026 funding round and valuation: https://www.channelnewsasia.com/business/ai-startup-genspark-valued-26-billion-in-latest-funding-round-6190486
Independent hands-on review after three weeks of use: https://www.lindy.ai/blog/genspark-review
Pricing analysis including the per-task credit estimates: https://www.eesel.ai/blog/genspark-ai-pricing
Vendor-reported benchmark coverage and the tool's architecture description: https://venturebeat.com/business/gensparks-super-agent-ups-the-ante-in-the-general-ai-agent-race
Review platform sources, all read 2026-09-28
Google Play listing and reviews: https://play.google.com/store/apps/details?id=ai.mainfunc.genspark
Apple App Store listing and rating: https://apps.apple.com/app/id6739554054
Trustpilot profile and review stream: https://www.trustpilot.com/review/genspark.ai
Product Hunt, workspace launch: https://www.producthunt.com/products/genspark
Community and community-reported evidence
Community subreddit, which carries the billing and session-limit discussion: https://www.reddit.com/r/genspark_ai/
The vendor's own account on the X platform, cited above for release announcements: https://x.com/genspark_ai
Framework and method
The U365 CI-First Evaluation Framework v1.2, including clause 5.2.3-a on agent-authored procedural memory, clause 4.2-a on agent-mediated conversation, and clause 7.5 on team-level rooms, which are the scoring rules applied in this review: https://www.university-365.com/ci-first
Published U365 INSIDE Tools reviews, read as comparisons and calibration for this series: https://www.university-365.com/tools
Vendor benchmark figures are quoted in this review as vendor self-reports and are labelled as such wherever they appear. No figure in this review is presented as an independent measurement of the agent surfaces.
Faculty Note on Evidence Quality
Four points about the evidence behind this review, each of which a reader should weigh before acting on it.
1. The one benchmark number attached to this product is a self-report from before the current product existed. The widely repeated figure in circulation is a benchmark score the vendor announced in 2025 for the first generation of its agent, relayed through technology press that labelled it as the vendor's claim. It is not on the official leaderboard for that benchmark, and it describes an architecture that predates the memory layer, the office suites and the team rooms reviewed here. This review records it as a self-report and does not treat it as evidence about the current product.
2. Nothing published measures whether the agent's work is correct, and that absence is the single most important fact in this review. No task-completion rate, no hallucination rate, no accuracy figure on any vendor surface; no independent evaluation of the agent surfaces from any source read for this review; and no per-run cost figure for the headline capability either. The Quality sub-score of 4 records that absence rather than an observed defect, which is the same treatment this series gives every tool in a category that ships no measurement.
3. The two pricing surfaces disagree with each other, and one of them is not readable without an account. The business pricing card and the help centre print different credit allocations for the same Team and Enterprise plans, both read on the same day, and the main pricing page redirects to a sign-in. Every published price in this review therefore carries a caveat: the figures are the vendor's own entry points, and the full credit ladder is visible only inside the product.
4. The community evidence is genuinely split, and the split is informative rather than contradictory. The mobile listings carry roughly 100,000 reviews at 4.5 and describe a large, mostly satisfied base using the product's simpler surfaces. The paying-customer review sites carry a few hundred reviews, moved recently from near 1.5 to 3.8, and describe a smaller cohort angry about the meter. This review treats the second group as the more informative one for a buyer, because it is the group that has paid and kept using the product, and because its complaint, the meter, is the same finding the framework reached from the scoring side. Where a reader finds a claim in this review that rests on a vendor page, the page is named; where a figure is a third-party estimate, it is labelled as an estimate.
Status and Re-check
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 published quality measurement of any kind. The vendor publishes no accuracy, hallucination or task-completion figure for the agent surfaces, and no independent measurement of their reliability exists. Any figure with a stated method, or an entry on an official agent benchmark leaderboard, would require the Quality sub-score to be re-run.
A change to the credit model, the session limits or the fair-use window. Credits do not roll over, an upgrade that changes the billing interval revokes unused credits, and users report a recurring session limit that resets on a fixed window. Any of those moving would change the Time sub-score and the cost advice in this review.
A change to the promotional position. Core chat and image generation are included at no credit cost, and the vendor states that the arrangement for members who subscribed before 18 September 2026 runs unchanged through 31 December 2026, after which the plan terms in effect at that time apply. The end of that arrangement is a re-check.
A memory control appearing on the SecondBrain or Claw surface. Neither surface offers a per-write approval step, and neither offers a routine practice of reading what was written. A review surface, a write-approval gate, or an export of the compiled context would change the clause 5.2.3-a outcome and the Skill sub-score.
A change to the training position for individual plans, or to the data-processing terms. The business pages state that customer content is never used to train models, and the team and enterprise help pages state that both paid team tiers are automatically opted out. The individual plan position is not stated on the same surfaces. A published position for individual accounts is a re-check.
A change to the Microsoft 365 or Google Workspace integration, or to the region and residency options. The plug-ins and the connectors are part of the product's stated value, and residency is offered on the enterprise tier only.
Enforcement or rule-making that lands on autonomous agents acting in a person's own accounts. The product sends mail in the user's voice, joins shared channels under the user's accounts, and runs scheduled tasks unattended. Any regulator or platform acting on that class of behaviour is a re-check.
The refund and cancellation terms changing. The refund window is 24 hours on monthly plans and 72 hours on annual plans for most customers, 14 days for the European Union, the United Kingdom and Turkey, and 7 days for Korea, with a single-refund rule and a rejection clause for heavy use before the request.
Comments