Two core members of Claude Code depart; Microsoft and Meta simultaneously scale back internal Claude usage

By | Qidian Pai

On October 24, within 45 minutes, two of Claude Code's most prominent members successively announced their departures on X.

One was Thariq Shihipar, a core engineer on the Claude Code team, whom the developer community calls the "King of Explanations," and who is connected to both the skills system and context engineering. The other was Anthony Morris, the lead creator of the Claude Code desktop version, who said his last day at Anthropic was October 23.

Neither said where they were headed next, and both farewell notes were written with grace. Shihipar even left a kind word for his former employer, saying his colleagues were the most talented group of people he had ever met.

That day was only about three weeks away from Anthropic's planned IPO. The company's target valuation had been pitched as high as $2 trillion, with the potential to set a global IPO record. Before the roadshow even began, the two most famous faces on the product line had already left.

01 Two big spenders hit the brakes at the same time

The departures are only the visible thread. Earlier, on October 5, The Information reported something else: Anthropic's two largest enterprise customers, Microsoft and Meta, were both cutting back internal employee use of Claude.

Earlier this year, Microsoft executives estimated that the company's annual internal spending on Anthropic technology would reach at least $1 billion. That budget has now been cut by more than a third, and employees have been asked to use Claude less and switch to GitHub Copilot and in-house models.

The cuts are even more granular when it comes to quotas. Microsoft's Cloud and AI division has more than 60,000 employees, and previously each person had a monthly AI model budget cap of $100,000; recently, most scenarios have been compressed to about $10,000, a 90% cut. Some engineers have complained that they can no longer use models whenever they want.

On Meta's side, the number of employees using Claude Code internally has fallen from about 60,000 earlier this year to about 30,000. Part of that halving is the aftermath of spring layoffs, when Meta cut about 10% of its positions. But the bigger reason is that Meta is pushing its own tools: the internal-only Meta Code already has more than 30,000 users, and more than 6,000 employees are using the externally tested Muse Code.

One number is worth breaking down. Even after the user base was halved, Meta still spent more than $105 million on Claude Code in the most recent 28 days. Half the people left, but the bill barely slimmed down.

What Microsoft cut was internal office and R&D spending. The portion Microsoft procures from Claude for external customers for Copilot features, originally expected to be at least $2 billion a year, is reportedly still rising steadily. Likewise, Meta and Microsoft still sell Claude to external customers through their respective cloud platforms.

So the accurate statement is: the big spenders did not cancel Claude; they simply stopped letting their own employees use it without limits.

02 Being expensive has become a sin

Why the contraction? Looking back at Microsoft's timeline, the answer is a bit brutally simple: Claude Code was too good to use.

In December 2025, Microsoft enabled Claude Code for thousands of engineers in its Experiences and Devices division. No one promoted it, no one mobilized for it; it took off by word of mouth, and token consumption doubled directly in the first few months.

The problem then became an arithmetic exercise. Each engineer's monthly token cost was between $500 and $2,000, and for a team combined that meant millions of dollars. For the first time, AI tools, like cloud resources, needed budgets, audits, and quotas.

Behind this is Claude's price position. Based on official pricing as of October 2026, the main Opus models cost $4 to $5 per million input tokens and $20 to $25 per million output tokens, with the flagship tier at $10 and $50, placing it overall on the most expensive side of the first tier. Its model capabilities have long topped the rankings, and its prices have long topped them too.

For the past two years, enterprises buying large models looked at performance leaderboards: whoever was strongest, they bought. Now procurement departments have changed their logic: whoever offers the best token value for money stays. OpenAI's GPT series has cut prices repeatedly this year, and Microsoft's internal Copilot has defaulted complex queries to OpenAI models since July, with the reason stated plainly: improving the cost-effectiveness of token investment.

03 Cracks in the two-trillion valuation

Put these events back on Anthropic's timetable, and their weight becomes clear.

According to media reports, Anthropic plans to complete its IPO as early as mid-November, launch its roadshow in the week of November 9, and aim to list before Thanksgiving. Around October 22, there was also news that the company could publicly file its S-1 as early as that week, with the window left to it narrowing.

Its financial cards have gradually been revealed through draft offering materials. In the six months ended June 30, the company's revenue was about $16.3 billion, compared with only $1.1 billion a year earlier; annualized revenue had already exceeded $65 billion in August, and the market expects it to surpass $110 billion by year-end.

At the same time, its net loss in the first half of the year was $80.7 billion, and its planned future computing power spending is as high as hundreds of billions of dollars.

The most glaring line is the customer structure: in 2025, nearly a quarter of Anthropic's revenue came from two customers. The media did not name them, but as its two largest enterprise customers, Microsoft and Meta are hard not to match to that description. The risk factors in the prospectus also include a sentence: among the largest group of customers, many have not signed long-term contracts and may reduce or even stop spending.

Now that sentence is becoming reality. The two biggest spenders are contracting internal usage in the same month, and the market's reaction is to directly adjust expectations: some investors have already lowered Anthropic's fair valuation from about $2 trillion to about $1.5 trillion. The AI sector has also been selling off over the same period, with CoreWeave down about 30% in two weeks.

Anthropic's situation can be summed up as a mismatch. Revenue growth is real, more than tenfold in half a year; customer loyalty, however, is conditional, and the condition is price. Its growth engine is the massive consumption of a handful of giant customers, and those customers happen to be the companies globally most capable of building their own alternatives and most skilled at doing the math. Big tech pulling AI capabilities back in-house is a move that has already begun and is hard to reverse.

There are several dates worth watching in the next three weeks: how detailed customer concentration is in the public S-1 filing, how management answers questions about Microsoft and Meta during the roadshow, and whether it can list before Thanksgiving as hoped. If the listing goes smoothly, whether Claude is expensive will be a matter of Wall Street voting with real money; if it is delayed, the brakes applied by these two big spenders today will be written into every analysis of the reasons for the delay.

For domestic large-model companies, the reference value of this is also direct. Anthropic has already demonstrated where the ceiling lies for the path of driving revenue high through a small number of major customers. Whether pricing can keep up with capability and whether the customer structure can be diversified are two questions that must be answered sooner or later; going public merely turns the answers into public information.