> 📥 The full analysis is available as a PDF (Danish, 6 pages). Download it at the top of the page. Round 1 of 3 in the series "Cowork vs. ChatGPT Work".
I spent 20 percent of a month's AI usage in a single day.
It was my first real workday with ChatGPT Work, OpenAI's new agent, launched on July 9. I gave it the same kind of tasks I give Claude Cowork every day: read the folder, gather research, finish the document.
It delivered fine. But when I opened the usage dashboard that evening, a fifth of the month's included agent usage was gone. At that pace the allowance is empty after five workdays. The rest of the month has to be bought as extra credits, and what those end up costing, I honestly don't know yet.
A lot of people are doing that same math right now. And even more haven't discovered that Work and Cowork exist at all. They keep working in the chat, unaware that both vendors now sell something closer to a digital colleague than a chatbot.
So let's take it from the top: what are the two products, what do they actually cost, and how do they compare on data protection?
1. Two agents, same promise
Claude Cowork started as Anthropic's desktop agent in January and rolled out to web and mobile on July 7. It gets access to your folders, your tools and your skills, works independently for long stretches and delivers finished files: documents, spreadsheets, presentations. Two days later OpenAI launched ChatGPT Work, widely described as the answer to Cowork.
ChatGPT Work does broadly the same thing. You give it a goal, it gathers context from your apps and files, works for hours and delivers finished spreadsheets, slides, documents and even websites. At the same time OpenAI is consolidating everything into one new desktop app where Chat, Work and Codex live together.
Functionally the two are converging. The difference sits somewhere else: in the business model.
2. Anthropic sells capacity. OpenAI sells a taximeter
With Claude you buy fixed capacity. Pro at 20 dollars a month (17 billed annually) includes Cowork, Claude Code and the rest. Max gives you 5x or 20x the Pro usage for 100 or 200 dollars. The Team plan starts at 2 seats today: Standard seats at 25 dollars per month (20 annually) and Premium seats with 5x Standard usage at 125 dollars (100 annually). Seat types can be mixed freely.
Hit the ceiling and you can turn on usage credits and pay API rates for the overflow. But the point is that a Premium seat or a Max plan is dimensioned for the agent to work all day. The bill is known up front.
OpenAI looks different. The ChatGPT Business seat costs 25 dollars a month (20 annually, minimum 2 seats), and ordinary chat is effectively unlimited. That sounds cheap. It is — right up until the agents start working.
Work, Codex and the Excel agent share one included usage allowance per seat, and it sits at the level of a 20-dollar Plus subscription. After that the meter runs: credits, billed per token. A typical agent task costs 5 to 45 credits, and OpenAI's own rule of thumb for Codex users is 100 to 200 dollars per month per developer at full throttle. Independent estimates land at roughly 4 cents per credit.
> List price vs. real price. 20 % of the month's included agent allowance spent in one workday. Five days at that pace and it's empty. 100-200 dollars extra per month at full throttle, per OpenAI's own rule of thumb. Fixed capacity gives a known bill; the credit model swings with usage.
Then there's the reset joker. My dashboard shows a monthly limit that resets, plus one available Full reset: a free reset of the entire usage limit, which OpenAI introduced for Codex in June and now also applies to Work usage. Run dry early and you pull the joker for a fresh allowance. Generous, yes. And also a sign of how new all of this is: the documentation still describes usage as 5-hour windows and weekly caps, while the dashboard shows a monthly allowance. The two stories don't quite meet yet.
If you want heavy agent usage at OpenAI without the taximeter feeling, you have to move up to ChatGPT Pro at 100 or 200 dollars. Those are consumer plans, and why that's a problem I'll get back to shortly.
3. My own math
My Claude setup is a Team plan with one Premium seat. That gives me Max-level usage for Cowork, every day, at a fixed price, with business terms on top.
My ChatGPT Business plan spent 20 percent of the month's agent allowance on a single day of Work, and after the first few days the meter is already at a third. Keep that pace and I can pull my one free reset and then buy credits for the rest. With OpenAI's own average numbers, that's 100 to 200 dollars extra per month, on top of the seat.
That doesn't make ChatGPT Work a bad product. It means the list price and the real price are two different numbers, and the gap grows with exactly the behavior the product is built for: letting the agent work a lot.
4. DPA and data: closer than most think
The good news: on data protection itself the two business plans resemble each other more than the debate suggests. Neither trains on your data by default. Both come with SSO, central administration and a data processing agreement (DPA). The difference is friction.
At Anthropic the DPA is built into the commercial terms that Team and Enterprise plans run on. Nothing to negotiate or sign separately.
At OpenAI you have to request the DPA yourself and sign it electronically. I did it for TryZone's ChatGPT Business workspace: request sent, Ironclad signature, executed the same evening with OpenAI Ireland as the contracting party and standard contractual clauses (SCCs) for transfers outside the EEA. It works fine. But you have to take the initiative.
> Key point. Neither business plan trains on your data by default, and both have a DPA. The difference is friction. On consumer plans you're a consumer customer — and your conversations may feed model training unless you turn it off yourself.
And the thing many people miss: on consumer plans you are a consumer customer. At both vendors your conversations may by default feed model training unless you opt out. Claude Pro and Max run opt-out, and ChatGPT Plus and Pro follow the same principle. If you work with customer or personal data, that work belongs on a business plan with a DPA. Whichever vendor you pick.
EU data residency, incidentally, doesn't exist on either of the two small business plans. At OpenAI data residency is an Enterprise matter (ten regions), and at Anthropic the extended data controls, such as custom retention, also start at Enterprise.
5. So what should you choose?
My short answer, the same one I give clients, rests on three questions.
1. Will the agent work for hours every day? Then fixed capacity beats a taximeter. At Claude that means Max for daily personal use, and Team or Enterprise when company or client data is in play.
2. Is there customer data in the tasks? Then consumer plans are out, and the DPA must be in place before the agent gets access to anything.
3. Can the budget live with variable costs? If no, OpenAI's credit model is hard to plan around right now. If yes, the Business seat plus credits can be fine for occasional agent work.
And whichever vendor: open your usage dashboard tonight. It's the fastest way to find out which side of the math you're on.
Sources
All prices and figures are based on the following sources (USD, excl. VAT, status July 2026):
- Claude pricing (Pro, Max, Team, Enterprise): claude.com/pricing
- ChatGPT Business pricing and seat types: openai.com/business/chatgpt-pricing · help.openai.com/en/articles/8792536
- Flexible pricing / credits (Business, Enterprise, Edu): help.openai.com/en/articles/11487671
- ChatGPT rate card and Codex rate card: help.openai.com/en/articles/11481834 · help.openai.com/en/articles/20001106
- Codex/Work usage limits and average cost: developers.openai.com/codex/pricing · resets: developers.openai.com/codex/changelog (June 2026)
- Credits across Codex, ChatGPT Work and Excel: help.openai.com/en/articles/12642688
- Anthropic's DPA (built into the Commercial Terms): support.claude.com/en/articles/7996862
- ChatGPT Work launch July 9, 2026: Bloomberg, The Next Web, Forbes · Cowork on web and mobile July 7, 2026: TechCrunch, VentureBeat
> ⚠️ Disclaimer. General practical guidance — not legal or compliance advice. Always involve your own compliance or legal function and your DPO, and verify dates and status against primary sources. All prices are in USD excl. VAT and may have changed; always check the vendors' own pricing pages. Status as of July 2026.
📬 Subscribe to the newsletter "AI, Built Human" on Substack — weekly insights on AI in practice.
Stefano Vincenti · GenAI strategist and architect · External lecturer, IT University of Copenhagen & DIS Copenhagen · Cofounder & CTO BotTellMe · Partner, TryZone · aitrainer.dk