Two weeks ago we looked at the new Copilot, where chat stays on the seat licence and Cowork, Code, Autopilot and the frontier models go on a usage meter. The question we left open was what Microsoft would do about the cost of the models underneath. On 5 October, The Information gave part of the answer, and it concerns Microsoft's own bill before anyone else's.
The article is paywalled, so everything that follows from it is as relayed by Channel Insider, DigitalToday and MLQ. Microsoft executives reportedly projected earlier this year that the company would spend at least $1bn a year on Anthropic technology for its own staff.
That has reportedly been cut by more than a third after staff were told to use less Claude and more of Microsoft's own tools, and monthly AI spending caps in the Cloud and AI division reportedly fell from $100,000 per person to roughly $10,000 in most cases. Meta's internal Claude Code users have reportedly halved, from about 60,000 earlier in the year to around 30,000.
If your project teams run on Microsoft 365, the model answering your Copilot prompt is a cost line Microsoft controls, and Microsoft has just shown how it behaves when that line gets large. Two days later Anthropic cut the price of its smallest model by around 75%. The two moves belong in the same story.
What Microsoft did to itself
The internal cut is not a divorce. A Microsoft spokesperson confirmed, per DigitalToday, that the default internal coding model has switched to OpenAI and that engineers can still choose Claude, which also remains available to customers through Microsoft Foundry and inside Copilot.
Channel Insider adds that Microsoft restricted staff use of a newer Claude model while its lawyers reviewed Anthropic's data retention terms, the same issue that led Nvidia, Palantir and Booz Allen to pull back on Fable, which we covered on 21 September.
The intent has been public since June. Microsoft AI chief executive Mustafa Suleyman said this in a Bloomberg interview, as reported by The Decoder.
We pay a lot of money to Anthropic—so our goal is to reduce and ultimately eliminate that cost.
For a Copilot customer, that is a supplier's margin plan. Microsoft can lower what it pays Anthropic by routing requests elsewhere, and you will not see the swap unless you test output quality yourself.
Why the customer side is the bit to watch
The Information followed up on 6 October with a newsletter titled 'Can Microsoft Help Customers Cut Back on Claude?'. According to MLQ's summary, Microsoft's payments to Anthropic for customers' use of Claude in Copilot have been rising and roughly offset the internal reduction. Microsoft is reportedly testing cheaper OpenAI and in-house MAI models for some Copilot tasks and trimming multi-step workflows so they burn fewer frontier tokens, in one reported case by reusing a standard PowerPoint script instead of regenerating it each time.
All these don't sound unreasonable. Bloomberg reported in July, via The Next Web, that Microsoft had already begun routing some app features to its own MAI models, while OpenAI and Anthropic still handled most Copilot traffic. The Decoder's reading is that cheaper MAI models could become the default with third-party models as premium add-ons. That is an inference, and Microsoft has not announced it.
Satya Nadella set out the right response at the Copilot launch, in a line GeekWire reported.
My test of true independence going forward is for you to have a private eval that you can hill climb on and substitute a model.
If the vendor swaps models underneath you, your own test tasks are the only defence.
Anthropic’s counter-move: price
On 7 October, Anthropic released Claude Haiku 5.5, with lower prices at the centre of the announcement. The changes cover API costs, cached inputs and subscriber credits.
Haiku gets much cheaper. Anthropic says the model is around 75% cheaper than Haiku 4.5 on average. The detail matters: prices are down 90% for requests up to 100,000 tokens and 50% for larger requests. Anthropic says 90% of Haiku 4.5 requests fall into the cheaper band.
The savings are clear in the API rates. For requests under 100,000 tokens, input costs fall from $1 to $0.10 per million tokens, while output costs drop from $5 to $0.50. Both are 90% reductions.
Cached inputs cost less, too. Sonnet 5.5 cache reads are now $0.10 per million tokens, half the previous price. Anthropic estimates this could cut costs by around 20% for most agentic workloads.
Subscribers get API credits. Max and Team subscribers receive between $100 and $500 a month in API credits. Haiku 5.5 is also available on Azure from launch day, giving enterprise customers another route to access it.
The wider context is cost pressure. We cannot say Microsoft's recent price cut prompted Anthropic's move. Neither company has said that. But the timing points to a market where the cost of running AI is becoming harder to ignore.
Subscription plans may offer unusually good value, for now. SemiAnalysis tested subscription limits on 5 October. Its detailed tables are paywalled, but the reported finding was that Anthropic's $200 plan gives heavy agent users around five times the API-equivalent value of OpenAI's plan at the same price. Treat that comparison with caution: subscription plans can subsidise usage to attract customers, and today's value does not guarantee tomorrow's pricing.
With widespread usage comes the need to manage efficiency and cost and to find the best model to suit the task at hand.
This is a customer quote on a vendor page, so it deserves that context. It also reflects the cost challenge Microsoft has just addressed internally.
Takeaway
The firms in this story set per-person spending caps, changed the default model and kept the expensive one available on request. A consultancy or contractor on Microsoft 365 can copy that pattern in a week, and should do it before the metered Copilot features reach the tenant. Keep a short set of your own deliverables, say a cost plan check and a variation assessment, and rerun them whenever the model behind a tool changes.
Ask your Microsoft account team which model handles each Copilot task today and how you would know if that changed. Decide which work needs a frontier model and route the rest to the Haiku price band. Treat the seat price as the floor, because the meter is where the money will go.
Project Flux will report what the model swaps inside Copilot and Claude actually do to output quality and cost for project teams. Subscribe at projectflux.ai to get those tests before your next licence renewal.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
