In May 2026, Microsoft began cancelling most internal Claude Code licences for its Experiences + Devices division, the group responsible for Windows, Microsoft 365, Outlook, Teams, and Surface. Engineers were instructed to transition their workflows to GitHub Copilot CLI by 30 June, which is also the last day of Microsoft’s fiscal year.
The instruction came from Rajesh Jha, Microsoft’s Executive Vice President for the division. Microsoft had opened Claude Code access to Experiences + Devices in December 2025, inviting thousands of developers, programme managers, and designers to use it in real engineering workflows. According to reporting by The Verge, the tool became popular quickly, popular enough that it started displacing GitHub Copilot CLI in daily use. That created a problem that was both financial and strategic: Microsoft sells Copilot to the rest of the world and cannot credibly do that while its own engineers migrate away from it at scale.
Jha’s internal message to staff, quoted by The Verge, was careful not to frame the decision as a reversal. “Claude Code was an important part of that learning,” he wrote. “At the same time, Copilot CLI has given us something especially important: a product we can help shape directly with GitHub for Microsoft’s repos, workflows, security expectations, and engineering needs.” The official framing is toolchain unification. The timing, the fiscal year boundary, and the cost pressure in the wider market suggest cost reduction was also a factor, though Microsoft has not said so directly.

The Uber figure, and what it does and does not mean

The Microsoft decision did not happen in isolation. It followed a disclosure from Uber that has been circulating in enterprise technology circles since April: Uber’s CTO for Mobility and Delivery, Praveen Neppalli Naga, confirmed to The Information that the company had exhausted its entire 2026 AI coding tools budget by April, four months into the fiscal year.

The mechanism was not a single large contract. It was adoption velocity. Uber rolled out Claude Code and Cursor to its engineering organisation in December 2025. The company introduced an internal leaderboard ranking teams by total AI tool usage volume, which accelerated uptake of both tools sharply. By March, 84 per cent of Uber’s roughly 5,000 engineers were classified as agentic coding users. Monthly per-engineer costs reached between $500 and $2,000 for heavy users, against an average of $150 to $250 across the organisation. At that rate of consumption, a full-year budget evaporated before summer.

COO Andrew Macdonald, speaking on the Rapid Response podcast and reported by Fortune on 26 May, gave the most direct public account of the resulting discomfort. “That link is not there yet,” he said, referring to the connection between rising AI spend and consumer feature output. “Maybe implicitly there’s more that is getting shipped, but it’s very hard to draw a line between one of those stats and ‘Okay now we’re actually producing like 25% more useful consumer features.'” Uber had told engineers that AI tool usage was good, incentivised it with a visibility mechanism, and then discovered that good incentives produce exactly the behaviour they reward, regardless of whether the underlying ROI case has been established.

The $3.4 billion figure that has appeared in several secondary reports refers to Uber’s total research and development budget for 2025, not its AI coding tools budget specifically. The figure that was exhausted was the AI tools allocation within that envelope. The distinction matters: the Uber situation is a story about AI tool spend outpacing its own budget line, not about Uber spending $3.4 billion on Claude Code.

Why token-based billing changes the forecasting problem

Both the Microsoft and Uber cases reflect a structural difficulty that a number of enterprises are encountering for the first time. Traditional software licences are predictable: a fixed number of seats at a fixed price. Agentic AI tools priced on token consumption are not. Usage scales with the ambition of the task. An engineer running a multi-step refactor across a large codebase generates vastly more tokens than one asking for a function suggestion. When agentic workflows become standard and developers run tasks in parallel, the per-user cost envelope widens in ways that quarterly budget models built in late 2025 could not easily have anticipated.

Anthropic moved Claude Code from flat-fee pricing to usage-based billing for autonomous agents earlier this year, a shift that reflects the same economic reality from the other side: agents use far more compute per task than standard chat interactions, and flat fees priced for chat do not sustain the infrastructure cost of agentic use at enterprise scale. Fortune cited a Gartner projection in its 26 May report that while per-token inference costs will fall roughly 90 per cent by 2030, enterprise AI bills will not fall proportionally because agentic workflows require far more tokens per task, and because AI providers will not fully pass cost reductions through to customers.

The result is a market where usage-based pricing is rational for vendors, adoption incentives inside organisations have a documented tendency to outrun the budgets attached to them, and the tools are genuinely useful enough that restricting access after a blowout feels punitive to the engineers affected.

Where Opus 4.8 enters the picture

Anthropic launched Claude Opus 4.8 on 28 May 2026, three days after Fortune published the Uber story and roughly two weeks into the Microsoft transition. The timing is coincidental rather than causal, but the proximity is not irrelevant to how the launch has been received.

The release made Opus 4.8 available immediately via the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and GitHub Copilot. That last integration is directly relevant to the Microsoft story: engineers in Experiences + Devices who are being moved from Claude Code to GitHub Copilot CLI now have access to Opus 4.8 through the tool they are being directed toward. The model is available to Copilot Pro+, Business, and Enterprise users, with a 15-times premium request multiplier applied until Copilot’s usage-based billing launches on 1 June.

On pricing, Anthropic held standard API rates flat at $5 per million input tokens and $25 per million output tokens, matching Opus 4.7. The new Fast mode, available to organisations in research preview via their account manager, is priced at $10/$50 per million, roughly three times cheaper than the Fast mode pricing on Opus 4.7. Anthropic described the release as a “modest but tangible improvement” over Opus 4.7, with benchmark gains in agentic coding, reasoning, and knowledge work. The most specific capability claim in Anthropic’s announcement is that Opus 4.8 is approximately four times less likely than Opus 4.7 to let flaws in code it has written pass without flagging them, a reliability characteristic that matters specifically in agentic workflows where the model is reviewing its own output.

It would be a misreading to treat the flat pricing as a direct response to the Uber and Microsoft situations. Anthropic has held Opus pricing steady across several recent releases. But the context in which enterprise buyers are evaluating Opus 4.8 is one where the cost structure of agentic Claude deployments has become a live CFO-level concern at two of the most prominent adopters on record. The launch landed into that conversation whether Anthropic intended it to or not.

What the Microsoft move actually settles, and what it does not

The clearest reading of the Microsoft decision is that it is about platform control as much as cost. GitHub has deep integration with code review, pull requests, CI pipelines, and repository metadata. Microsoft owns GitHub. Copilot CLI built on top of that integration has structural advantages for internal Microsoft workflows that Claude Code, sitting outside that ecosystem, cannot easily replicate. Jha’s reference to Microsoft’s “repos, workflows, security expectations, and engineering needs” is an accurate description of why a vertically integrated tool wins on governance grounds even when it loses on developer preference.

What the decision does not settle is whether Claude Code is less capable than Copilot CLI for the engineering tasks Microsoft’s teams actually perform. The internal signals point the other way: engineers adopted Claude Code at a rate that required an active policy decision to reverse. Copilot CLI now inherits a comparison it did not ask for, and engineers who preferred Claude Code will be evaluating it against a tool they were using by choice.

Uber’s situation is different in kind. Uber is not pulling back on Claude Code for platform reasons. It ran out of budget because adoption worked precisely as intended, and because the ROI case for agentic coding spend has not yet been made in terms that connect to consumer product output. CTO Neppalli Naga described the company as “back to the drawing board” on AI budgeting. Uber has also said it plans to test OpenAI’s Codex alongside Claude Code as it explores broader agentic deployment, which suggests the budget problem is being treated as a forecasting failure rather than a signal to stop.

The broader pattern here is not that enterprise AI coding tools are failing. It is that organisations incentivised adoption without building the financial and governance infrastructure to manage what successful adoption looks like at scale. That is a solvable problem, and both Uber and Microsoft appear to be solving it, in different ways and for different reasons.

The primary sources for this piece are: The Verge’s reporting on the Microsoft internal decision, Fortune’s 26 May 2026 article by Jake Angelo on Uber’s COO comments, The Information’s earlier reporting on the Uber budget blowout cited by Fortune, and GitHub’s 28 May changelog entry on Opus 4.8 availability in Copilot.