ChatGPT Work and Claude Cowork now do the same job. Choose on metering.

Yes, within about eight weeks, and most published comparisons have not caught up.

ChatGPT Work launched on 9 July 2026 as OpenAI's agent for knowledge work. It runs on the same execution and isolation machinery as Codex, inside an isolated virtual machine, with a managed Chrome instance the agent controls through tool calls. It produces finished artifacts rather than chat: spreadsheets, documents, slides, and hosted Sites. Rollout reached Pro, Enterprise and Edu first and Plus and Business within days. Reporting put it past ten million users within three weeks, though that figure is OpenAI's own and covers Work and Codex together, so treat it as a scale signal rather than an ops adoption number.

Claude Cowork launched in January 2026 as a desktop agent: give it a goal and it works across your files and tools. It runs on macOS, Windows and Linux, reaches local files and folders, connects to services including Microsoft 365, Google Drive and Slack, and carries its own built in browser so it can open sites and fill forms without borrowing yours. Plugins bundle skills, connectors and sub agents into one install.

The convergence happened from both directions at once. OpenAI put a desktop mode with real local access on a product that began as a cloud sandbox. Anthropic put web and mobile on a product that began on the desktop, explicitly so people could supervise long running tasks away from the machine running them. The result is that the sentence "one is cloud, one is local" is now false about both.

This matters beyond trivia, because that sentence was doing all the work in most buying decisions. Take it away and a lot of the confident advice written about these two products in the middle of the year is now advice about a distinction that no longer exists. If you shortlisted one of them in the spring on that basis, the shortlist is worth redoing.

Neither vendor will tell you, and that is the single most important fact in this comparison.

ChatGPT Work has no fixed task quota and no per task price. Every task consumes a variable slice of your plan's allowance depending on how much gathering, reasoning and building the agent decides it needs, so the same request can cost more on Tuesday than it did on Monday. Claude is structured similarly: usage limits apply, the Max tiers are sold as five times or twenty times more usage than Pro, and the actual thresholds are not published on the pricing page.

So both products are sold on a seat price and metered on something the buyer cannot see. That has three practical consequences, and they apply equally to both:

• You cannot forecast next month from this month. A workflow tool that bills per run lets you multiply. An agent that decides its own step count does not. If your plan for this technology requires a number in a budget, the number you can defend is the seat cost, and nothing below it.
• The pilot is systematically cheaper than production. A trial runs a handful of considered tasks. Production runs the same task hundreds of times with nobody watching. Teams read a comfortable pilot as evidence the economics work, when all it proved is that the economics work at pilot volume.
• The hedge is a paid tier, not a better estimate. This is why both vendors sell a high usage escape hatch: ChatGPT's Premium seat, introduced in August 2026 at $100 per seat billed annually, and Claude's Max tiers from $100 a month. Notice that those are the prices at which unpredictability stops being your problem and starts being theirs. That symmetry tells you the metering opacity is a deliberate design on both sides, not an oversight on one.

The way to handle this is not to pick the cheaper sticker. It is to run four weeks on whichever tool you already pay for, on real work, and measure what your own team actually consumes before you commit to an annual seat count. That is the only number that will be true for you, and it costs you nothing to collect because you are already paying for the seat.

Run down these in order and stop at the first one that matches. They are ordered by how much money the answer saves you.

• You already pay for one of them across the team. Use that one. The capability gap does not repay a migration, and the real cost of switching is not the subscription, it is retraining habits and rebuilding connector access. Stop here.
• You want to test agentic ops work before asking for budget. Claude Cowork, on a Pro seat at $17 a month billed annually. It is the cheapest legitimate way to find out whether this changes anything for your team, and you are not blocked from the core product by a plan gate.
• You need shared agents governed at the organisation level. ChatGPT Work, on Business. Workspace agents and company knowledge are the pieces that turn individual experiments into something an operations lead can administer, and they are Business and Enterprise only, so a Plus seat will not show you what you are buying.
• Your stack is long tailed. ChatGPT Work. More than a thousand plugins covers more obscure SaaS than a curated connector set will. Check your three most awkward tools against both directories before you decide, because this is the one axis where a single missing integration settles it.
• The work has to be auditable and must not fail silently. Neither, on its own. A general agent that goes wrong returns a confident, plausible, wrong result and no error. If a task repeats on a schedule and someone downstream depends on it, it belongs in a workflow with an error path, with a model called at the one or two steps that genuinely need judgement. We made that case in full in buy the agent or build the workflow.

The last one is the one teams skip, and it is the one that costs the most. Both of these products are excellent at work that arrives differently every time and ends in a document a person reviews. Neither is a substitute for a workflow you can monitor. If your shortlist came out of a search for an agent, it is worth checking that the job you had in mind is actually agent shaped before you compare seats at all. If it is not, that is our AI Engineering stage rather than a subscription decision.