ChatGPT Work vs Claude Cowork: Which AI Agent Wins?

Forty-eight hours. That’s the gap between Anthropic pushing Claude Cowork onto the web and phones (July 7) and OpenAI launching ChatGPT Work alongside GPT-5.6 (July 9). The two biggest AI labs shipped the same product category in the same week, and neither of them is pretending it’s a coincidence.

So now there are two serious “hand the AI a whole project” tools, and everyone deciding between them is stuck comparing launch posts written by marketing teams. This is the comparison those posts won’t give you — including the numbers where each side loses.

One honest note up front: this is week one for ChatGPT Work and week one of Cowork’s mobile era. Details below will age. The shape of the choice won’t.

The 30-second verdict

If you just wanted the answer: free or budget-first → ChatGPT Work. Your job lives in Microsoft 365 → Cowork. You want an artifact factory → Work. You want a doer with hands → Cowork. The rest of this piece is why — and where each answer breaks.

What each one actually is

ChatGPT Work is the agent workspace OpenAI shipped with GPT-5.6: brief it on a project, it works in the background — minutes to hours — and returns a finished file. Spreadsheets with real formulas, formatted decks, long documents, even small working web apps. It runs on the same infrastructure as Codex, OpenAI’s coding agent, pointed at office work. We wrote a full beginner guide if you’re starting from zero.

Claude Cowork is Anthropic’s version, and it’s been maturing since January — which shows. As of July 7 it runs on desktop, web, and phones; keeps working in the cloud after you close your laptop; runs on a schedule if you ask; and — the headline — its Microsoft 365 write tools let it send email, manage your calendar, sort your inbox, and edit SharePoint files. Our how-to-use guide covers the basics.

Same category. Genuinely different philosophy — and that’s the part worth understanding before any spec sheet.

The real split: makers vs. doers

Wait — that deserves plainer language, because it’s the entire decision.

When ChatGPT Work finishes, you get a thing. The worst realistic failure is a wrong thing, which you catch (or don’t) when you open it. Nothing has left the building. When Cowork finishes, something may have happened — a message sent from your address, a meeting on someone’s calendar, a file changed where your team works. Most of Cowork’s actions ask permission along the way, but the Microsoft 365 connector’s send permission is literally called Mail.Send, and the documentation makes no promise of a final “are you sure?” before a message leaves your mailbox.

Neither philosophy is “right.” An agent that can’t touch your real tools makes you the courier for everything it produces. An agent that can touch them makes your attention the only safety net. Pick based on which failure mode you’d rather manage — not which demo looks cooler.

Price and access: one of these is free

ChatGPT Work is included on every plan — including free. Free and Go users get GPT-5.6 Terra inside Work (a model that beats last generation’s flagship on several benchmarks), paid plans pick between Sol, Terra and Luna, and Pro/Enterprise unlock ultra, which throws four parallel agents at one task. That free tier isn’t a crippled demo; it’s a real agent with OpenAI’s newest mid-size model. Full plan mapping here.

Cowork has no free door. It needs a paid Claude plan — it works on Pro at $20/month, but there’s a catch that’s easy to miss: the new capabilities roll out to the $100+ Max plans first. The July 7 mobile and background features went Max-first, with other plans promised “in the coming weeks.” On Claude, the shiny thing in the announcement is usually not the thing your $20 plan has today.

Platforms cut the other way: Cowork is on your phone; ChatGPT Work isn’t yet. Cowork sessions keep running in the cloud and ping your phone when Claude needs a decision. If your workday happens between meetings rather than at a desk, that’s not a small difference. One asterisk: mobile and web Cowork can’t reach your local files the way desktop Cowork can — same name, less reach.

Models and muscle: the benchmark question

Under the hood it’s GPT-5.6 versus the Claude 5 family, and for once we have both vendors’ numbers plus a neutral referee.

From OpenAI’s own launch table — which, credit where due, includes rows OpenAI loses:

And the referee: Cursor — the AI code editor, no allegiance to either lab — runs CursorBench on real user sessions: ambiguous, multi-file, badly-specified tasks. The opposite of contest conditions. Current standings: Claude Fable 5 Max on top at 70.5%, GPT-5.6 Sol Max just behind at 67.2% — but Sol’s run cost $5.22 to Fable’s $17.32. Claude keeps a small, real quality lead; OpenAI wins on price by roughly 3×.

Now the disclaimer, and read it twice, because benchmark tables decide more purchases than they should. Vendor numbers come from vendor launch posts — they choose the tests, the competitors, and the settings (that 91.9% is ultra, a four-agents-at-once mode you won’t run casually). Benchmarks are contest conditions: clean inputs, defined goals, no office politics, no “actually the deadline moved.” Your work is ambiguous, half-documented, and full of context no leaderboard sees. A three-point gap on a chart says nothing about which agent handles your client’s tone, your spreadsheet conventions, your inbox. The scores above tell you both tools are frontier-grade and neither sweeps the board. They cannot tell you which one is better for you — only a week of your own tasks can.

The bill nobody mentions in the demos

Both agents share an uncomfortable truth: agent work burns usage allowance at a rate chat never did.

On the ChatGPT side, Work shares rate limits with Codex, and launch-week users watched single substantial tasks eat most of a usage window — ultra multiplies the burn by design. On the Claude side, one user reported a single Cowork task — building a presentation — consuming roughly $45 of usage, and a Pro subscriber found one five-hour Cowork session ate about 15% of a weekly allowance. Anthropic doubled Cowork’s five-hour limit through August 5, but weekly caps didn’t move: you burn faster, you don’t get more.

The practical rule is identical on both: treat agent runs like meetings, not messages. Schedule the ones that earn their cost. Write the brief properly the first time — five sloppy runs cost five times one good one.

Trust, safety, and who’s being careful

Anthropic’s pitch has always been the careful one, and the Cowork permission model mostly earns it: Claude borrows your Microsoft credentials rather than getting its own, can only see what you can see, sends are attributed, an org admin has to consent before write tools exist at all. That’s a genuinely sound design — with the one gap flagged above (no guaranteed confirmation on connector email sends) and the standing risk every inbox-connected agent carries: prompt injection through inbound mail.

OpenAI’s answer is architectural rather than procedural: ChatGPT Work mostly doesn’t act in your systems — it makes things and hands them to you. Fewer permissions to audit because there’s less it can touch. You are the deployment mechanism, which is slower and safer.

For a solo user, honestly, either posture is manageable. For an IT admin, the questions differ: with Cowork you’re auditing permissions (Mail.Send, Calendars.ReadWrite); with ChatGPT Work you’re auditing what employees paste into it and where the artifacts go. Pick your paperwork.

So which one, for you?

You’re on a budget or just curious → ChatGPT Work, and it’s not close. Free access to a frontier agent with no card is this launch’s genuine gift. Learn agent-briefing on Terra; decide about paying later.

Your job runs on Outlook, Teams and SharePoint → Cowork is playing a game ChatGPT Work hasn’t entered: it can actually do things in Microsoft 365, not just draft them. Budget for Max if you want the new features now, and pilot the write tools with training wheels on.

You produce documents, decks and analyses all day → ChatGPT Work’s artifact factory is purpose-built for you, and free Terra means you can test that claim today.

You’re a developer → different calculus: Claude still leads real software-engineering benchmarks by a wide margin (SWE-Bench Pro: 80.3 vs 64.6), while OpenAI’s ultra mode and pricing are aggressive. Run both against your repo; also see the Copilot Cowork comparison if you’re in the Microsoft stack.

You manage a team and have to pick ONE → start with what the team already pays for. These tools reward existing context (connected accounts, learned habits) far more than benchmark deltas. The switching cost is real; the three-point score gap isn’t.

You genuinely can’t decide → run the same three real tasks through both this week. One will need less babysitting on your work. That’s your answer, and no blog — including this one — can compute it for you.

What neither of them will do

• Know your standards. Both produce competent-generic; the last 20% that makes work yours stays yours.
• Check their own facts. Both will format a wrong number beautifully. Verify anything that feeds a decision.
• Stay still long enough for this comparison to be final. OpenAI shipped free-tier agents; Anthropic will answer. Cowork got write tools; expect ChatGPT Work to grow hands eventually. Revisit in a quarter.
• Replace knowing the work. You can’t review a financial model you couldn’t have sketched. Agents raise the value of judgment; they don’t retire it.

The bottom line

ChatGPT Work versus Claude Cowork isn’t really OpenAI versus Anthropic — it’s maker versus doer. One builds you artifacts and leaves the acting to you, free tier included. The other reaches into your actual tools, sends your actual email, and asks you to manage the trust that requires, at a paid-only price. The benchmarks say both are frontier-grade and neither sweeps; the referee eval says Claude’s a little better and OpenAI’s a lot cheaper; and none of those numbers knows your job.

The skill that transfers across both — and across whatever ships next quarter — is directing an AI agent well: briefing, reviewing, correcting. That’s exactly what our ChatGPT vs Claude course builds head-to-head, and ChatGPT Workspace Agents for Non-Engineers goes deep on the no-code agent side. Learn to delegate; the logo on the agent matters less every month.

Sources

• ChatGPT is now a partner for your most ambitious work — OpenAI (July 9, 2026)
• GPT-5.6: Frontier intelligence that scales with your ambition — OpenAI (July 9, 2026)
• Claude Cowork on web and mobile: hand off work anywhere — Anthropic (July 7, 2026)
• Anthropic brings Claude Cowork to mobile and web — VentureBeat
• CursorBench: model evals on real Cursor sessions — Cursor
• OpenAI releases GPT-5.6 and ChatGPT Work tool — Axios