Microsoft, Uber tell employees to stop using Anthropic's Claude amid surging AI compute costs

Microsoft and Uber are being cited as early examples of a problem many large companies expected to solve later: generative AI usage is becoming a line item big enough to force product and procurement decisions. IBTimes UK reported that both companies urged some employees to stop using Anthropic’s Claude as bills rose, reframing AI adoption from an experimentation story into a cost-control story. Anthropic’s own pricing pages show why that pressure can build quickly on coding-heavy workloads, especially when premium models are used at scale. The immediate point is not that companies are abandoning AI tools. The point is that usage is no longer abstract. Once internal assistants move from pilots to daily engineering workflows, token consumption, caching strategy, model selection and fallback rules become finance questions as much as technical ones. That is the backdrop for the reported pullback around Claude use at Microsoft and Uber. (platform.claude.com) ### Why would Claude usage become expensive fast inside large companies? Anthropic’s current API pricing shows a steep spread between model tiers. Claude Opus 4.7 starts at $5 per million input tokens and $25 per million output tokens, while Claude Sonnet 4.6 is priced at $3 and $15, and Claude Haiku 4.5 at $1 and $5. Anthropic also says prompt caching can cut costs materially, but only if workloads are designed to use it well. (platform.claude.com) Coding workloads are a particularly costly case because they often involve long contexts, repeated repo scans, multiple tool calls and iterative output. Anthropic markets Opus 4.7 for “professional software engineering” and “complex agentic workflows,” and says the model can run long tasks in the background. Those are exactly the use cases that can drive heavy token usage inside a large engineering organization. (platform.claude.com) ### What, specifically, has been reported about Microsoft and Uber? IBTimes UK reported that Microsoft and Uber were rethinking AI rollouts as computing costs came in higher than expected, including reports that some workers were told to stop using Claude. Separate secondary reporting, citing The Verge, said Microsoft had begun canceling thousands of internal Claude Code licenses and was steering developers toward GitHub Copilot CLI, with a reported June 30, 2026 cutoff for much of the usage inside its Experiences + Devices group. (anthropic.com) For Uber, Forbes reported that the company burned through its 2026 AI budget in four months on Claude Code. Yahoo Finance’s pickup of the Forbes report said Uber had also ranked engineers on internal leaderboards tied to Claude Code usage, which the report said encouraged heavier consumption. Reuters could not independently verify those internal claims from the available public documents reviewed here. (letsdatascience.com) ### Why does this change how companies build AI products? Anthropic’s documentation already reflects a portfolio approach rather than a one-model strategy. Its pricing and product pages separate premium Opus models from cheaper Sonnet and Haiku options, and highlight caching and batch processing as cost levers. That tends to push companies toward routing systems that decide when a task truly needs a frontier model and when a smaller model will do. (forbes.com) In practice, that means more work on orchestration layers: prompt caching, model routing, usage caps, approval gates and small-model fallbacks. Anthropic’s consumer and enterprise pricing pages also show explicit spend controls and usage analytics, a sign that cost governance is becoming part of the product surface rather than only a back-office concern. ### Does this mean vendors that manage AI spend may benefit? (platform.claude.com) Anthropic’s own pricing structure suggests there is room for software that helps companies avoid overpaying for premium inference. If a company can shift a share of requests from Opus to Sonnet or Haiku, or improve cache-hit rates, the savings can be immediate because the token prices differ sharply by tier. That makes spend-control, observability and routing tools easier to justify in procurement terms. (claude.com) The next public marker to watch is June 30, 2026. That is the reported date by which Microsoft’s Experiences + Devices group is expected to wind down most Claude Code use, according to reporting that cited The Verge, while Anthropic’s pricing pages remain the clearest public reference for how companies are likely to measure the trade-offs. (letsdatascience.com) (platform.claude.com)