
Copilot, ChatGPT or Claude: Microsoft's Cost Study Is Really About the Meter
On 16 November Microsoft published a cost comparison that will end up in a lot of 2027 budget meetings. In its post What AI at work really costs, the company says that for an "everyday AI user" Microsoft Copilot costs $30 a month, ChatGPT Enterprise $81 and Claude Enterprise $205. For an "advanced" user who also hands work to agents, the figures are $96, $149 and $299. Quality, according to the same study, was broadly comparable across the three.
Those numbers will be quoted as "Copilot is the cheapest". The more useful thing in the study is the explanation of why the numbers differ. All three products now charge, at least in part, for how much people actually use them, and each one puts the meter in a different place. That, more than the price of a seat, is what decides your bill.
Who measured what
The study was commissioned and funded by Microsoft. Microsoft ran the testing and the cost analysis in October 2026; GigaOm contributed to the design, validated the method and results and co-wrote the findings. The full report, Demystifying the TCO of Enterprise AI, is on GigaOm's portal.
The setup was narrow and Microsoft says so. Testers used 37 prompts across Copilot Chat, Cowork and Copilot in Word, Excel and PowerPoint, scaled to a modelled month of 100 tasks for an everyday user and 140 for an advanced user, 20 of which were agentic workflows. All three products worked against the same fictional company and its data, which lived in Microsoft 365. ChatGPT and Claude reached that data through connectors and add-ins. Prices are October 2026 list prices, or assumed prices where none are published, without negotiated discounts. Microsoft's own conclusion is to "read the findings as directional".
None of this makes the study worthless. It makes it a study of one particular situation: a company whose documents, mail and meetings are already in Microsoft 365, buying at list price.
Where each meter sits
Microsoft's explanation of the gap is that everyday chat and Office usage is included in the Copilot user subscription, while agentic workflows run on usage-based Copilot Credits. In the study, competitors' subscriptions may be $10 cheaper than Copilot, but "once AI usage charges were added, that advantage disappeared within the first week of modeled use."
The rival products publish their own billing rules, and they really are built differently.
For Claude Enterprise, Anthropic's billing page is blunt: the seat fee gives access to Claude, Claude Code and Cowork "but doesn't include any usage. Every token your team consumes is billed separately at standard API rates." There is no individual token allowance per seat. Administrators can set spending limits for the whole organisation and for individual users. Self-serve Enterprise plans are billed in US dollars only; another currency requires a sales-assisted contract.
For ChatGPT Enterprise, OpenAI's flexible pricing page says core features such as search, file upload and canvas need no credits. Advanced features (Deep Research, Thinking models, image generation, advanced voice, Codex) draw from a shared credit pool bought at contract level. There are no per-seat caps by default, and on credit-based agreements the workspace overage limit "starts at No limit", so usage can continue after the pool runs out and be billed under the agreement unless an owner sets a cap.
So the three vendors charge for different things: Microsoft for agents beyond the subscription, OpenAI for advanced features beyond the pool, Anthropic for every token. Which one is cheapest for a given company depends on what its people actually do all day, and on the mix of light and heavy users.
The gap narrows as the work gets agentic
There is a detail in Microsoft's own figures that gets lost in the headline. For everyday users, Claude comes out at almost seven times the Copilot cost ($205 against $30) and ChatGPT at 2.7 times ($81 against $30). For advanced users the ratios fall to about 3.1 and 1.6 ($299 against $96, $149 against $96).
My reading is that this is the more durable finding. Once work moves from chat to agents, every vendor charges by consumption, Microsoft included, and the advantage of a bundled subscription shrinks. If your 2027 plan is mostly about agents and automations rather than chat in Word and Outlook, the comparison that matters is cost per completed task, and the study's own numbers suggest the differences there are smaller than the headline.
The quality method is worth copying regardless of which tool you buy. Before testing, evaluators wrote a checklist for each prompt (right sources, right people, right figures, complete deliverable, no unrequested actions), and a run passed only if it met every item. Partly correct work counted as a failure. That is a stricter standard than most internal pilots apply.
A hypothetical test before you sign
Take a hypothetical distribution company with 250 employees, most of its work in Microsoft 365, and a renewal discussion coming up in the first quarter. Three departments want different tools.
Rather than adopting someone else's average user, the company could pick 15 to 20 real tasks from those departments: a weekly sales summary, a supplier complaint reply, a price-list reconciliation in Excel, one multi-step agent task. Each task gets a written pass checklist in the same style as the study. Each is run several times in each candidate tool over two to four weeks, with the vendor's real offer rather than list price. Alongside the quality score, the team records what the usage meter showed: credits drawn, tokens billed, agent runs.
At the end the company has a cost per passed task for its own work, split by the kind of user who does it. It may still choose Copilot, or decide that ten heavy analysts need a different tool from everyone else. The decision rests on measured work rather than on a modelled month built by one of the bidders.
Two settings belong in the test from day one. In ChatGPT Enterprise, set an overage limit rather than leaving the default. In Claude Enterprise, set organisation and per-user spend limits. Gartner recommends the same discipline for agents, including session-level spending caps under what it calls agentic FinOps, in a release published the same day as Microsoft's study.
What to do with the study now
Use it as a template, not a verdict. Microsoft itself says the better test is your own, and offers the methodology and full prompt list through its account teams. Ask for it, adapt the prompts to your company, and run the same test on every product on your shortlist, Gemini included if you are a Google Workspace shop, since it was not part of this comparison.
Then make someone responsible for the meter after the purchase. Usage-based AI spending does not stay where the pilot left it: it grows with adoption, and it changes when vendors retire models or adjust rates. Monitoring that spend alongside quality is ordinary AI operations work, and deciding which tool gets which budget is a leadership call that needs a named owner, whether that is a CIO, a finance lead or an AI director, in-house or as a service.
Microsoft Ignite opens today, 17 November, and the session list includes several talks on AI spend management. Licensing details may shift again this week. The question for your budget stays the same: what does a completed piece of your work cost, in each tool, at your volume?


