Internal use of Anthropic's AI products is falling sharply at two of its biggest corporate customers, with Microsoft cutting its Claude-related spending by more than a third and Meta Platforms, Inc. (Meta) seeing its employee user base for Anthropic's coding assistant roughly halve. The pullback comes as both companies push workers toward their own in-house models and tools.

Meta and Microsoft are Anthropic's two largest enterprise customers, and their belt-tightening poses a challenge to the startup's ambition of sustaining rapid growth. Two people familiar with the spending said Microsoft executives estimated earlier this year that internal use of Anthropic technology would cost at least $1 billion annually. One of the people said Microsoft has since cut that budget by more than a third, asking employees to limit Claude usage to save money and lean more on Microsoft's own AI tools. The figure excludes fees generated by Microsoft platform customers calling Anthropic models, a revenue stream that is still climbing steadily.

At Meta, the number of employees using Anthropic's coding assistant Claude Code has dropped sharply to around 30,000, according to people familiar with the matter. Internal Meta data reviewed by The Information showed the figure stood at about 60,000 earlier this year. Part of the decline stems from Meta's previous round of layoffs, which eliminated roughly 10% of its 78,000 employees. But the bigger driver is Meta's continued development of its own AI coding tools and its Muse Spark family of large models, which are sold externally and also mandated for priority internal use.

Meta's homegrown answer to Claude Code, called Muse Code and built on Meta models, opened to outside customer testing in August, and internal data shows more than 6,000 employees have recently been using it. A separate internal-only tool, MetaCode, has more than 30,000 users. Meta's internal AI tool budget had originally been expected to reach several billion dollars this year. An internal memo did not specify how that budget would be split among outside vendors, though Claude Code was the most popular external coding tool at the time. Internal data shows Meta spent more than $105 million on Claude Code in one recent 28-day period. In developing its consumer-facing Muse agent this year, Meta initially relied on Claude, but the final product launched last month switched to its own models.

Anthropic faces mounting growth headwinds. Although the two tech giants remain major Claude customers, even modest spending cuts by top accounts can put broad pressure on Anthropic. Surging AI tool costs are prompting other enterprise customers, such as Cognition, to seek cheaper alternatives. While few companies can build their own replacement models the way Meta and Microsoft can, many are turning to open-source models and rival offerings from OpenAI and Google, whose aggressive pricing keeps squeezing Anthropic's room on price. Some of that migration is lifting OpenAI's revenue, which has surged about 70% on an annualized basis in recent months to near $70 billion.

Citing data-privacy concerns, large institutions that handle sensitive enterprise work, such as Palantir Technologies Inc. (Palantir), Nvidia and Booz Allen Hamilton, are either demanding new data-security guarantees from Anthropic or reducing or dropping use of its most advanced and expensive models. These challenges arise as Anthropic prepares for an IPO that could be one of the largest ever, requiring the company to disclose more business details to potential investors. Reuters, citing draft IPO prospectus data, reported that Anthropic's revenue last year was highly dependent on a small number of large customers. Revenue has surged this year: annualized revenue reportedly topped $65 billion in July, more than seven times higher than at the start of the year, alongside a wave of new customers.

Anthropic has told investors that roughly 6,000 customers spent at least $100,000 on its AI software in the 12 months through June, four times the number for all of 2025. Over the same period, more than 1,000 enterprise customers spent over $1 million each, and more than 100 spent over $10 million each.

Meta and Microsoft are both customers and competitors of Anthropic. Meta Chief Executive Mark Zuckerberg announced last week the creation of a Meta Enterprise Platform division dedicated to selling AI products to business customers. Meanwhile, Anthropic's rapid revenue growth and flood of new enterprise products caught Microsoft management off guard over the past year. This spring, Claude usage inside Meta surged as employees competed on an internal "AI Token Legend" leaderboard to burn through tokens. That prompted management in June to tighten AI usage controls, set budgets and token limits, and launch a central monitoring dashboard to track excessive use of external AI tools such as Claude Code.

Microsoft employees' use of Claude also ballooned earlier this year. Two people familiar with the spending said that, as of May, Microsoft's internal annual outlay on Claude models, including Claude Code, Claude embedded in Microsoft Copilot, and the Mythos model used for cybersecurity scanning, was on track to exceed $1 billion. Rapidly rising Claude bills among engineers, combined with warnings about GitHub losing its edge in AI, led several executives to ask employees to rely less on Anthropic products. Management cut Claude Code subscriptions and instead directed employees to call Claude models through GitHub Copilot. Copilot lets users manually select a model, but defaults to OpenAI models in automatic mode. For Microsoft, that approach costs far less because, under its commercial partnership with OpenAI, Microsoft pays nothing extra to use OpenAI models.

People familiar with the matter said Microsoft executives including Azure cloud chief Scott Guthrie and developer tools head Jay Parikh have asked employees to use coding models prudently, urging department leaders to monitor staff AI spending and prioritize non-Anthropic models. A current Microsoft employee said executives can now see each employee's daily AI spending and the specific models they call on GitHub Copilot. An internal memo Parikh sent employees in August, reviewed by The Information, said: "Migrating more workloads internally to OpenAI models helps improve the value of token spending. This fiscal year, every department has a token budget, and everyone needs to keep learning and optimizing usage." (404 Media previously reported parts of Parikh's directive.)

Microsoft has since further lowered overall AI usage limits for employees. For example, the cloud and AI division, which has more than 60,000 employees, previously had a monthly AI model budget cap of $100,000 per person; the cloud unit recently cut the monthly budget for the vast majority of employees to about $10,000. Current employees say the lower limits have frustrated some engineers who complain they cannot use models freely. Microsoft is cutting internal spending on Anthropic, an amount that excludes what Microsoft pays Anthropic to deliver Claude features to external customers, which is significantly larger. Microsoft had originally expected to spend at least $2 billion a year buying Anthropic models for AI features in its Copilot office software for outside customers. Microsoft is also switching models in Copilot; people familiar with the matter said that starting in July, complex Copilot queries default to OpenAI's GPT-5.6. But one person familiar with the matter said that, as Copilot subscription users grow, Microsoft's payments to Anthropic for external services are still rising steadily this year. That increase largely offsets the reduction in Microsoft's internal Claude spending, leaving total payments to Anthropic roughly flat versus earlier highs this year.

A Microsoft spokesperson confirmed the company recently set OpenAI models as the default for internal AI coding, though engineers can still manually select Claude. The spokesperson said Microsoft steers employees toward its own GitHub Copilot coding tool rather than external tools like Claude Code, because internal use of its own products helps iterate and refine them. In a statement, the Microsoft spokesperson said: "Microsoft's business groups need to manage AI usage prudently and efficiently in line with business priorities and standard engineering and operational practices."