On 25 September Microsoft kept the $30 seat and put Cowork, Code, Autopilot and the newest models on a credit meter above it, without saying what a task costs. Priced against ChatGPT and Claude Enterprise at 500 seats, Copilot is the dearest option only for light and idle users. At medium and heavy use it is the cheaper one, and the direct vendors still trail it on compliance. Twenty-seven pages on what the seat buys now and what the meter leaves unpriced.
On 25 September 2026 Microsoft announced a redesigned Copilot app with three sections, Home, Code and Autopilot, and a companion post moved the new work onto Copilot Credits at $0.01 each. Chat stays inside the $30 seat. Cowork, Code, Autopilot, new SharePoint features and frontier models such as Fable and Astra are metered on top. Much of the commentary read this as Microsoft dropping the seat and folding GitHub Copilot into one product. Neither happened. The seat is still listed at $30.00 per user per month, GitHub Copilot stays a separate subscription, and what converged is the unit of account: GitHub AI Credits and Copilot Credits both cost a cent.
That leaves a finance lead who can cap what Copilot spends but cannot forecast it. Microsoft publishes the price of a credit and a full rate card for Copilot Studio agents, and no credit amount for a Cowork task, a Code app or an Autopilot hour. It points buyers to an estimator instead. Two defaults also lean toward spending: new Copilot Business licences bought through resellers arrive with pay-as-you-go switched on from 2 November 2026, and an "Auto-apply new services" setting adds whatever Microsoft ships next to existing spending policies.
The report tests the argument that usually follows: that Copilot is the most expensive way to buy AI, that a firm could go direct to OpenAI or Anthropic and connect it to Microsoft 365 for less with equal admin controls, and that Microsoft failed by never building its own model. It prices a 500-seat firm on list prices at three levels of use, compares admin controls from each vendor's own documentation, and reads Microsoft's filings on what it actually pays for models. The verdict is half right. Microsoft is the dearest option for light and idle users, and the cheaper one once people use it.
Every claim below carries an evidence label in the report itself: fact, vendor claim, third-party estimate, assessment, hypothesis or scenario. Every price is a US list price read off the vendor's own page on 26 September 2026, before any enterprise agreement discount. The cost figures are the author's own model and are labeled as assessments. Where a number could not be found, such as the seat price of ChatGPT Enterprise, the report says so rather than estimating it.
Microsoft's enterprise page still lists Microsoft 365 Copilot at $30.00 per user per month paid yearly, and the billing post says "there are no changes to the value you already have." Microsoft's own announcement says developers keep using GitHub Copilot, and GitHub is changing how its seat is billed, not ending it. What converged is the meter: GitHub AI Credits and Copilot Credits both cost $0.01.
Microsoft prices the credit and the Copilot Studio rate card, which charges 1 credit for a classic answer, 2 for a generative answer, 5 for an agent action and 10 for tenant graph grounding. It publishes nothing per Cowork task, per Code app or per Autopilot hour. For the two newest features, the work is where the money will go, and the work has no price.
On the author's model at 500 seats, Microsoft costs $30.00 per user per month against $22.52 for Claude Enterprise at light use, but $37.02 against $42.14 at medium use and $82.65 against $112.34 at heavy use. At a plausible mix of users the two land within $110 a year of each other across all 500 seats. ChatGPT Enterprise cannot be totalled because OpenAI does not publish its seat price.
SSO, SCIM, no-training defaults and hard spending caps are documented on all three. Native DLP, retention and eDiscovery are not. Microsoft Purview covers ChatGPT Enterprise for audit and eDiscovery but not for DLP or sensitivity labels, and Anthropic relies on third-party DLP partners such as Forcepoint and Cyberhaven and documents no EU data residency.
Anthropic's Microsoft 365 connector reads and, where an admin allows, writes across SharePoint, OneDrive, Outlook and Teams, and its Excel, PowerPoint and Word add-ins are generally available. OpenAI has live apps for SharePoint, Outlook and Teams and one add-in covering Word, Excel and PowerPoint, but no Outlook add-in. Both run through Microsoft Graph on permissions Microsoft sets.
Since April 2026 Microsoft no longer pays OpenAI a revenue share, while OpenAI keeps paying Microsoft one through 2030. Satya Nadella calls the OpenAI model licence royalty free, with rights running through 2032. The only reported outside model bill is Anthropic, on pace for about $500 million a year in January 2026 according to The Information, against $24.1 billion Microsoft booked from OpenAI arrangements in fiscal 2026.
Mustafa Suleyman says the OpenAI contract capped Microsoft's own frontier training until late 2025. Since then Microsoft AI has shipped seven MAI models, including one tuned for Excel, Nadella has tied model work to cost of goods, and every new Copilot feature sits on one credit meter. That is most of the strategy the argument says Microsoft missed. It was late, not wrong.
Microsoft reported over 30 million paid Copilot seats in July 2026 against over 450 million commercial Microsoft 365 seats, about 6.7 percent. Paid seats doubled between January and July. The weakness is after purchase: Recon Analytics found only 35.8 percent of workers with Copilot access use it, and where Copilot, ChatGPT and Gemini were all available, 8 percent chose Copilot and 70 percent ChatGPT.
Microsoft 365 Copilot and Claude Enterprise are the only two options fully priced at 500 seats, because ChatGPT Business stops at 200 employees, Claude Team at 150, and OpenAI lists ChatGPT Enterprise as custom pricing. Claude Enterprise is a $20 seat with no usage included, plus usage at API rates. Direct chat is priced on Sonnet 5 and agent tasks on Opus 5.5, with each agent task reading one million input tokens, half from cache, and writing 30,000 output tokens, which costs $2.70 at list. Microsoft's chat sits inside the seat, and its agent tasks are priced 35 percent below the direct cost, the midpoint of Microsoft's own claim that Cowork is 30 to 40 percent cheaper than Claude Cowork.
Ten chat turns a working day over 21 days, 4,000 tokens in and 400 out per turn, and no agent tasks. The direct usage bill is $2.52, so Claude Enterprise lands at $22.52 against the flat $30.00 Copilot seat. Across a year at 500 seats that is $180,000 against $135,120, a $44,880 premium.
$30.00 against $22.52
Thirty chat turns a day at 6,000 tokens in and 600 out, plus four agent tasks a month. The direct usage bill reaches $22.14, and the Copilot seat already covers the chat. A year at 500 seats is $222,120 on Microsoft against $252,840 on Claude Enterprise.
$37.02 against $42.14
Medium chat plus 30 agent tasks a month, a direct usage bill of $92.34. At 500 seats that is $495,900 a year on Microsoft against $674,040 on Claude Enterprise. The size of that gap rests on Microsoft's own unaudited claim about Cowork cost.
$82.65 against $112.34
Sixty percent light, thirty percent medium and ten percent heavy users. The annual totals at 500 seats come to $224,226 on Microsoft and $224,328 on Claude Enterprise. The argument that Microsoft is the most expensive way to buy AI holds only in the light-use row.
$37.37 against $37.39
If Microsoft's agent tasks cost exactly what the direct API costs, Copilot rises to $40.80 at medium use and $111.00 at heavy use. That is still below Claude Enterprise in both rows, by a few dollars. Microsoft is not the most expensive option at heavy use under either reading.
$40.80 and $111.00
Recon Analytics found only 35.8 percent of workers with Copilot access use it. Applied to 500 seats, that leaves about 321 idle seats, and at a $10 monthly gap between an idle Copilot seat and an idle Claude Enterprise seat, about $38,520 a year spent on nothing.
About $38,520 a year
The crossover is simple. Microsoft's seat costs $10 more than Claude's and includes chat that Claude bills by the token, so once a user's direct chat bill passes $10 a month, which on these token sizes happens somewhere between 26 and 40 chat turns a working day, the Copilot seat is the cheaper one before any agent work is counted. ChatGPT Enterprise cannot be placed on the chart. At OpenAI's published Enterprise token rates for GPT-5.6 Sol, the same light and medium chat profiles cost $5.04 and $22.68 per user per month, double the Sonnet 5 figures, before any seat fee. The $60 figure that circulates for a ChatGPT Enterprise seat traces to a single 2023 forum post and is not used. For firms of 300 or fewer the picture changes again: Copilot Business lists at $21.00, or $18.00 on a first-year promotion for annual commitments made before 31 December 2026, which makes it the cheapest AI seat of the three vendors. The model leaves out three costs that fall only on the direct route: third-party DLP tooling, connector administration and a second vendor contract.
The report compares what each vendor documents on its own pages, control by control. On identity, training terms and spending caps the direct vendors are level with Microsoft, and on caps OpenAI and Anthropic are cleaner. Microsoft's credit limits stop usage when a user hits them, but budgets on its older pay-as-you-go billing for Copilot Chat agents only send email, and Microsoft's documentation says that system "doesn't enforce the budget or prevent your organization from exceeding it."
Compliance is where going direct costs more than the invoice shows. A Microsoft shop gets Purview DLP, retention and eDiscovery for Copilot inside tools it already runs. Going direct means a Compliance API feed into a separate DLP product, and for ChatGPT, Microsoft's own Purview page says DLP and sensitivity labels are not supported. Anthropic documents US or global data residency only, with no EU option.
Microsoft's strongest counterargument, one vendor and data kept inside the tenant, is weaker than it sounds. Anthropic models are on by default for most commercial customers outside the EU, EFTA and UK, and they are excluded from the EU Data Boundary. Cowork does not show the model provider in audit records. And Microsoft's own documentation says SharePoint's default sharing setting is the most permissive option, so every route in this report, Copilot included, reads whatever your users can already open.
The cost ranking depends almost entirely on how much each person uses AI, so the first step is a usage census, not a vendor shortlist. Light users and non-users are cheapest on Copilot Chat, which Microsoft includes at no extra cost, or on a direct Enterprise seat. Heavy users are cheapest on the Copilot seat if Microsoft's Cowork claim holds, and close to parity if it does not. Treat compliance as the deciding test rather than a tiebreaker: if your DLP, retention and eDiscovery run on Purview and must cover AI prompts today, Copilot is the only option that does it natively.
Before the first spending policy goes live, turn off "Auto-apply new services" and set per-user credit limits. If you buy Copilot Business through a reseller after 2 November, check whether usage billing arrived switched on and at what limit. Do not approve Code or Autopilot beyond a pilot until Microsoft publishes what they consume. For the finance lead, the number to ask for is the share of seats in active use, not the seat count, and the budget needs two lines: a fixed seat line and a credit line with hard per-user caps, funded from an Azure consumption commitment if one exists.
The report closes with three scenarios to the end of 2027 and the earliest visible sign of each. Metered Copilot becoming the default buy sits at 45 percent, signalled by a published per-task credit table for Code or Autopilot or an MAI model in the Copilot model picker. A split estate, with Copilot for light users and ChatGPT or Claude Enterprise for the heavy minority, is 35 percent, and the tell is Purview DLP support for ChatGPT Enterprise or EU residency from Anthropic. The seat folding into E5 or E7 is 20 percent, and a cut to the $30 list price would be the first sign. The probabilities are the least reliable content in the report, and the tripwires are the part anyone can watch.
This isn't a vendor summary. Every sentence is labeled by what stands behind it: verified fact, vendor claim, third-party estimate, my assessment, hypothesis, or scenario. Sources are numbered and clickable. Forward-looking sections use scenarios with observable tripwires, not forecasts. It's the same method behind every market assessment I write.
Twenty-seven pages, built from public sources with no client brief and no interviews. Read it in the browser or take the PDF.
Each report here answers a real question, directed and researched against public sources and evaluated against a stated assumption, then delivered as Word and PDF. If you're weighing a platform, sizing a category, or defending a number to a board, tell me the decision behind it and I'll tell you honestly whether a report is the right tool.
Commission an assessment