Twice in one week, the two biggest AI labs answered the same question the same way. On July 7, Anthropic pushed Claude Cowork to the cloud, web, and mobile. On July 9, OpenAI shipped ChatGPT Work alongside GPT-5.6. Both pitch the identical thesis: an agent that takes a goal, works for hours, and hands you back a finished thing — a spreadsheet, a deck, a doc, an app — instead of a chat reply you still have to act on.

For a founder, the interesting question isn't which lab is smarter. It's which agent fits the shape of your actual work.

If you read one line: pick Cowork if your work lives in files on your machine; pick ChatGPT Work if your work lives in connected SaaS apps and you already spend your day inside ChatGPT.

They optimize for different surfaces#

This is the whole decision, so lead with it.

Cowork is a file agent. It reads, writes, and edits files directly on your Mac or Windows machine — no upload/download shuffle. It structures folders, builds presentations, and writes spreadsheets with working formulas, not screenshots of spreadsheets. It runs on Claude Opus 4.8 with a 1-million-token context window and coordinates sub-agents for parallel work. If your day is a research folder, a pricing model in a spreadsheet, or a codebase, Cowork is operating on your native material.

ChatGPT Work is an app agent. It's a mode that sits next to Chat inside ChatGPT. You hand it a goal; it gathers information across your connected apps and workflows, breaks the project into steps, and stays on it — returning finished sheets, slides, docs, or a shareable web app. It runs on GPT-5.6, OpenAI's July release, which ships in three tiers: Sol (top), Terra (mid), and Luna (fast and cheap). If your day is pulling context out of the SaaS tools you already live in, that's where ChatGPT Work reaches.

Same verb — do the work — aimed at two different filesystems. One is your disk; one is your app graph.

What each one actually returns#

The difference matters more than it sounds. A Cowork spreadsheet lands in the same folder your accountant's template lives in. A ChatGPT Work web app is a link you send. Neither is better; they end in different places, and the right one is wherever your next step already happens.

Both now run while your laptop is closed#

The quiet upgrade this month is that neither agent is chained to your session anymore.

Cowork's cloud remote sessions — beta, rolling out to Max plans first — run on Anthropic's servers, so a task keeps going after you shut the lid. Anthropic's own usage data, released with the cloud push, shows 91.3% of Cowork sessions have nothing to do with coding — a signal that "agent" has already outgrown "coding agent" for most users.

ChatGPT Work runs server-side too, and is built to stay with a complex project for hours, decomposing it and finishing the steps on its own. For a solo founder, this is the actual unlock: you can start a job, close the laptop, and collect the output later.

Connectors decide the real fit#

An agent is only as useful as what it can reach.

If your company runs on Microsoft 365, Cowork has a head start on your inbox and documents. If your stack is a spread of SaaS apps you've connected to ChatGPT, ChatGPT Work meets you there.

Price: you probably already own one#

Here's the part that ends most deliberations. Cowork is bundled free into every paid Claude plan — Pro ($20/mo), Team ($25/seat), Enterprise — with no separate SKU. ChatGPT Work is included in ChatGPT plans, rolling out to Pro, Enterprise, and Edu first, with Plus and Business following within days.

So for most founders, this isn't a buying decision at all. It's a which-subscription-do-I-already-pay-for decision. If you're on Claude, you have Cowork. If you're on ChatGPT, you have ChatGPT Work. The cost of trying both is one month of the other plan.

The verdict#

Neither product is the winner, because they aren't really competing for the same hour of your day.

The realistic outcome: you'll use whichever you were already paying for, and you'll be mostly happy. The one experiment worth running this month is pointing both at the same real task — a board-deck draft, a competitor teardown, a data cleanup — and seeing which one hands back something you'd actually ship. That answer is specific to your work, and it's the only benchmark that matters.

For the wider week — GPT-5.6's launch, MCP's stateless spec locking on July 28, and the rest of the agent wave — see our Founder's Wire. Building your own repeatable agent workflow? Start with an agent skill.