ChatGPT Work is OpenAI’s new agent, launched July 9, 2026 and powered by GPT-5.6, that takes a goal, breaks it into steps, pulls context from your files and connected apps, and returns finished documents, spreadsheets, or sites, pausing only when it needs a decision from you.
It ships today on Pro, Enterprise, and Edu plans, with Business and Plus rolling out within days. It sits between ChatGPT Chat (single-turn conversation) and Codex (software engineering), aimed at the long, multi-step projects that used to eat a Tuesday afternoon.
Here’s the contrarian part: ChatGPT Work is not a productivity upgrade. It’s a rental agreement. Every workflow you hand it lives on OpenAI’s infrastructure, under OpenAI’s pricing, subject to OpenAI’s next repricing cycle. You’re not building capability, you’re renting a very fast employee who forgets you the moment you stop paying.
What ChatGPT Work Actually Is
ChatGPT Work is an agentic mode inside ChatGPT that plans, executes, and completes multi-step projects using approved files, connected apps, and the web. It differs from ordinary ChatGPT chat in one way that matters: persistence.
Instead of answering a question, it holds a goal across hours, stops to ask for missing context, and returns a coordinated set of outputs rather than a single reply.
The desktop version, rolling out to Windows after launching on macOS, can also touch approved local files, applications, and browsers, not just cloud-connected tools. That’s a direct answer to Anthropic’s Cowork, which shipped the same local-file agentic pattern months earlier.
OpenAI built this for repeat, definable workflows: refreshing a course syllabus against accessibility guidance, assembling an accreditation packet, drafting a research brief with citations. None of these are one-off chats. All of them are the kind of recurring, multi-source project that previously required a coordinator, not a chatbot.
The catch is structural, not technical. ChatGPT Work is excellent at the task. It is silent on who owns the workflow once it’s built.
The GPT-5.6 Engine Underneath
GPT-5.6 ships in three variants, Sol for maximum capability, Luna for speed, and Terra as the balanced default, and ChatGPT Work routes between them depending on task complexity.
This tiered-model approach is now standard across frontier labs: use the expensive model only when the task demands it, because agentic workflows burn far more tokens than chat.
That token math matters more than the marketing does. An agent that runs for two hours autonomously isn’t answering one prompt, it’s executing dozens of internal reasoning steps, tool calls, and file reads. Gartner reports that 80% of enterprise applications shipped or updated in Q1 2026 embed at least one AI agent, up from 33% in 2024. Volume is exploding. So is the bill.
Rémy Bigot’s rule for any AI tool applies directly here: the capability is real, but capability without a cost ceiling is not a strategy.
ChatGPT Work vs Claude Cowork: The Real Difference
The market now has two credible agentic office tools, and they solved the problem differently. Claude Cowork runs primarily on your desktop, working your local file system directly, you point it at a folder and it executes there. ChatGPT Work runs primarily in the cloud, with the desktop app extending it to approved local files and applications as of this release.
The governance gap is the sharper distinction. OpenAI built Workspace Agents, the underlying architecture behind ChatGPT Work, for organization-wide sharing: build an agent once, deploy it across a team, centrally control which connectors each group can touch. Cowork, per current comparisons, gives admins spend and access controls but nothing close to that level of centralized agent governance yet.
Translation: OpenAI is betting on the enterprise IT department. Anthropic is still betting on the individual power user. Neither is betting on you owning the infrastructure either way.
We talk a lot about that on FutureRadar.
What ChatGPT Work Costs, in Real Numbers
ChatGPT Business runs $20/seat/month annual, or $25 monthly, two-seat minimum, per OpenAI’s April 2026 rate card. Enterprise pricing isn’t published, it’s negotiated, but disclosed buyer data puts it at $45–75 per user per month, with most contracts landing near $60, and a 150-seat minimum. That floors an Enterprise deployment at roughly $108,000 a year before Codex credits, API overages, or integration costs.
Run that math for a 20-person team on Enterprise-equivalent access and you’re at $1,000–1,500 a month minimum, and that’s before anyone builds a single custom workflow, because the workflows themselves aren’t yours. Cancel the subscription and every agent, every connector, every trained workflow disappears with it.
Where ChatGPT Work Wins
ChatGPT Work is the strongest choice when the task is genuinely cloud-native, cross-team, and needs centralized governance from day one. If your work already lives in Google Drive, Slack, and a dozen SaaS connectors, and your IT team needs to control who touches what, OpenAI’s Workspace Agents architecture is built for exactly that.
It also wins on breadth. 60+ connectors ship with the Business plan alone, and the « Sites » feature, turning a Work output into a shareable interactive page, is a genuine differentiator nothing else on the market matches yet. A program review that ends in a source-backed brief and a live internal tracker, updated automatically as source material changes, is a real capability upgrade over a static PDF.
None of that changes the ownership question. It changes who’s best positioned to rent from OpenAI specifically.
The Adoption Numbers Nobody’s Advertising
Only 31% of enterprises have even one AI agent in production, per S&P Global Market Intelligence and McKinsey, and just 23% of organizations report significant ROI from AI agents specifically, against 29% for generative AI broadly. Median time-to-value on an agent deployment is 5.1 months.
Those numbers should sit next to every ChatGPT Work headline. The tool works. The organizational discipline to make it pay back in under six months mostly doesn’t exist yet. 56% of enterprises now name a dedicated « AI agent owner, » up from 11% in 2024, which tells you the industry already knows agents without an owner become expensive toys.
Own vs. Rent: The Question ChatGPT Work Doesn’t Answer
Every agentic tool launch since Cowork has followed the same script: impressive demo, opaque enterprise pricing, workflows that live entirely inside someone else’s platform. ChatGPT Work is the best version of that script yet. It is still that script.
Terminator got one thing right that most AI coverage misses: the danger was never the machine doing the work, but who controlled it. A system that doesn’t tire, doesn’t churn, and executes exactly what it’s told is either your permanent advantage or your permanent dependency. Rent the agent, and you’re renting your own operating leverage from someone else, forever, at their next repriced rate card.
The logic scales down to any business, not just Fortune 500 IT departments running 150-seat Enterprise contracts. A managed AI agent stack built and owned outside the big-lab subscription model runs closer to $4,700/month fully loaded for a small team, fixed, predictable, and yours regardless of what OpenAI or Anthropic charge next quarter.
That’s the calculation worth running before signing anything with a 150-seat minimum: what does it cost to rent this capability forever, versus building it once. Agent Nexus (nexus.asymmetry-partners.com) exists for exactly that comparison.
ChatGPT Work vs Claude Cowork vs Owned Agent Infrastructure
| Factor | ChatGPT Work | Claude Cowork | Owned agent stack (e.g. Agent Nexus) |
|---|---|---|---|
| Execution environment | Cloud + approved local files | Local desktop-first | Your infrastructure |
| Governance model | Org-wide, centrally managed | Admin spend/access controls | Fully custom |
| Entry pricing | $20-25/seat (Business) | Included in Claude plans | ~$4,700/month flat, team-wide |
| Enterprise floor | ~$108,000/year, 150-seat minimum | No published minimum | No seat minimum |
| Workflow ownership | Rented, OpenAI infrastructure | Rented, Anthropic infrastructure | Owned outright |
| Best fit | Large orgs, cross-team, cloud-native | Individual power users, local files | Teams wanting fixed cost + ownership |
FAQ
Q: What is ChatGPT Work and how is it different from regular ChatGPT?
A: ChatGPT Work is an agentic mode that holds a goal across multiple steps, pulls context from files and connected apps, and returns finished outputs — rather than answering one prompt at a time like standard ChatGPT chat.
Q: Is ChatGPT Work available yet?
A: Yes, since July 9, 2026, on macOS desktop for all plans, with Windows and web/mobile rollout for Plus, Pro, Business, Enterprise, and Edu following within days.
Q: How much does ChatGPT Work cost for a business?
A: Business plan access starts at $20-25 per seat monthly with a two-seat minimum. Enterprise access typically runs $45-75 per user monthly with a 150-seat minimum, putting the annual floor near $108,000.
Q: Is ChatGPT Work better than Claude Cowork?
A: Neither is categorically better. ChatGPT Work has stronger centralized governance for large, cloud-native organizations. Claude Cowork works more directly with local files for individual power users. Both keep your workflows on the vendor’s infrastructure.
Q: Is ChatGPT Work actually worth it for a small business?
A: Rarely, at Enterprise pricing. Most small teams don’t clear the 150-seat minimum and don’t need org-wide governance. A Business-tier subscription or an owned agent stack usually beats renting Enterprise access nobody’s built to use yet.
Q: Why do most companies get AI agent adoption wrong?
A: They buy the subscription before naming an owner. Only 23% of organizations report significant ROI from AI agents, and the ones who do almost always have a dedicated agent owner and a defined workflow before deployment — not after.
Q: Does ChatGPT Work replace the need for custom AI agent infrastructure?
A: No. It replaces ad hoc chat sessions for defined, repeatable workflows. It does not give you ownership, cost predictability at scale, or infrastructure that survives a pricing change from OpenAI.
The Verdict
ChatGPT Work is the most capable agentic tool OpenAI has shipped, and it’s still a lease, not a foundation. If your workflows are genuinely cross-team and cloud-native, and you can absorb a $108,000 Enterprise floor without blinking, take it.
If you’re a growing business trying to build durable operating leverage instead of a recurring line item that scales with someone else’s roadmap, run the own-vs-rent math before you sign anything with a seat minimum.
Before your next tool decision, book 15 minutes at Asymmetriq and see what owning the infrastructure actually costs against renting it forever.