Is Microsoft Paying Anthropic $2 Billion a Year for Copilot? Claimed vs Verified

Microsoft is reportedly on track to pay Anthropic more than $2 billion a year so that Claude can power parts of Copilot, according to The Information, which published the claim on October 12, 2026. If accurate, that is about four times the roughly $500 million a year the same outlet reported in January. The catch: the figure comes from one unnamed source at one outlet, and neither company has confirmed it. This post separates what is reported from what is verified.

TL;DR: claimed vs verified

What exactly was reported

AI Weekly summarised The Information's report by Alexis Dufresne on October 12, 2026. According to that summary, a single source told the outlet that Microsoft is on track to pay Anthropic more than $2 billion a year for Copilot. The summary itself flags that the figure rests on one outlet and one source.

The same report says Charles Lamanna, an executive vice president who now leads Office, pushed CEO Satya Nadella to start paying for Anthropic models to power Copilot. He reportedly had to argue past internal worries about Anthropic's higher cost. The motive offered: customers complained Copilot was buggy and not good enough at automating work in Excel and PowerPoint, and Anthropic's latest models reportedly performed better on advanced Excel financial functions and PowerPoint generation.

Claude is described as being blended with OpenAI technology inside Word, Excel, Outlook and PowerPoint, with Microsoft still weighing both providers on capability and margins. We could not open the original article, which is paywalled, so everything here about its contents comes from the secondary summaries and should be treated that way.

Where the $500 million figure came from

The earlier number is also a report, not a disclosure. In January 2026, Sherwood News's Jon Keegan cited The Information as saying Microsoft was on track to spend $500 million per year on Anthropic's AI services. The same item said Microsoft lets salespeople count Anthropic sales toward their quotas, a rare move, and was using Anthropic models in GitHub Copilot and Microsoft 365 Copilot.

So we have two run-rate projections nine months apart, from the same outlet, both anonymously sourced. They tell a consistent story of rising spend. They do not give an audited year-over-year growth rate, and "on track to" can mean a forecast based on current usage, not a bill already paid.

What is actually confirmed

Some of the surrounding facts are on the record. In November 2025, Anthropic announced strategic partnerships with Microsoft and NVIDIA: Microsoft committed to invest up to $5 billion in Anthropic, and Anthropic committed to purchase $30 billion of Azure compute capacity. That is a two-way relationship: Microsoft is both an investor and a customer-and-supplier. It is also why a Copilot payment figure is hard to interpret from the outside, because money moves in more than one direction.

Also on record, from our own coverage: Microsoft and Anthropic have been tangled in billing questions before. A startup's Azure Foundry bill for Claude was passed between the two companies, which we covered in who bills whom on Claude in Azure Foundry. That shows Claude is already sold through Microsoft channels. It says nothing about the size of Microsoft's internal payments to Anthropic.

Timeline: how Microsoft's Anthropic relationship grew

Read the table top to bottom and the direction is clear, but only one row is a company announcement. The two reported rows are the ones carrying the dollar amounts.

How to read a single-source number

Big-dollar claims about AI lab deals spread quickly because they are easy to repeat and hard to disprove. A few habits help when you meet one:

• Ask what the unit is. Annual run-rate, committed contract value, and cash paid in a quarter are three different numbers. The phrase "on track to" usually means a projection from recent usage.
• Ask who benefits from the number being public. Anonymous sources inside a company can be accurate, but they sometimes have a stake in the internal debate. Here, the reported story is partly a story about one executive winning an argument.
• Look for the second source. A figure repeated by five aggregators is still one report. Count origins, not headlines.
• Check the denominator. Over $2B sounds enormous next to a single product, but Microsoft reported capital spending in the tens of billions and Anthropic's own compute commitments are far larger. As a share of either company's activity, it could be modest.

We applied the same lens in other claimed-versus-verified posts, such as the Anthropic Long-Term Benefit Trust reporting, where the facts that held up were the ones with documents behind them.

Why the number is hard to verify

• Microsoft does not itemise suppliers. Its filings report segment revenue and capital expenditure, not a per-model-provider cost line. Our post on the Maia 300 inference chip shows how the company talks about inference cost: in efficiency terms, not vendor invoices.
• Anthropic is private. Revenue figures reach the public through leaks and prospectus coverage, such as the numbers in the draft prospectus reporting. Even there, customer-level amounts are rarely broken out.
• "Pay" is ambiguous. Spending could include inference on Azure that Microsoft resells at a margin, direct API purchases, or committed capacity. Those are different things with different implications.
• Single-source reporting. One unnamed source can be right, but nobody else has corroborated it publicly.

Why this matters even if the exact figure is off

Even a smaller true number would signal something real: Microsoft decided Copilot needs a second model supplier for quality reasons, not just redundancy. That fits a pattern we have followed. Microsoft has been adding model choice to Office, including Grok models in Word, Excel and PowerPoint, and has been experimenting with enterprise Copilot discounts mixing seat and usage pricing. The strategic question is margin: every query routed to a pricier external model eats into Copilot gross margin, which is why the report mentions cost concerns.

It also lands in a week of Microsoft cost headlines, such as the voluntary exit offer to about 60,000 employees. Those are separate stories, and nothing we reviewed links them causally.

What this means for what you build or pay

• Do not anchor on $2B. Treat it as a rumor about direction, not a data point for modelling Anthropic's revenue or Microsoft's margins.
• Expect multi-model Copilot. If you build on Microsoft 365 Copilot extensibility, assume outputs may come from different underlying models over time. Test behaviour on your own tasks and pin versions where you can.
• Watch for price signals. If Microsoft's extra model cost is real, it may show up in Copilot packaging. Track Microsoft's pricing pages and earnings-call commentary rather than viral posts.
• Use model choice yourself. If Excel and PowerPoint automation quality is why Microsoft reportedly paid more, run your own evaluation on your spreadsheets before choosing a model.

What would count as confirmation

It helps to decide in advance what evidence would settle the question. A named executive at Microsoft or Anthropic giving a figure on the record would count, as would an earnings-call remark that quantifies third-party model spend. A second outlet with an independent source, not one repeating The Information, would also move the claim from reported to corroborated. A filing that describes the commercial terms would be stronger still. Until one of those appears, the sensible position is that the direction of travel is credible, the magnitude is unproven, and the contract structure is unknown. That is a narrower claim than the headline, but it is the one the evidence supports.

How to check this yourself

1. Find the original The Information article (October 12, 2026) and read who it attributes the claim to.
2. Look for Microsoft or Anthropic statements: newsroom posts, executive interviews, earnings call transcripts.
3. Search Microsoft's SEC filings for supplier concentration language. Absence of a line item is itself informative.
4. Be sceptical of any outlet that restates the figure as fact without a second source.

Bottom line

The claim is plausible and consistent with earlier reporting, but it is a single-source run-rate figure, unconfirmed by either company. We will update this post if Microsoft or Anthropic comment or a second outlet corroborates the number.

Related reading

• Claude on Azure Foundry: who bills whom
• Microsoft Copilot with Grok in Office
• Microsoft Copilot enterprise discounts
• Microsoft buyouts: verified and not
• Microsoft Maia 300 inference chip
• Anthropic draft prospectus numbers

Figures and reports are accurate as of October 27, 2026 and may change if Microsoft or Anthropic comment.