Meta spent more than $105 million on Claude Code during a single twenty-eight day period this year. The number appeared in internal reporting. Not long after, the company told its engineers to stop.
The headcount using Anthropic's coding assistant at Meta fell from approximately sixty thousand to thirty thousand, according to reporting by The Information published in the first week of October. At Microsoft, where the Cloud and AI division had permitted individual employees to consume up to $100,000 per month in AI model usage, the ceiling dropped to around $10,000. The company's earlier projection that it would spend at least $1 billion annually on Anthropic's technology internally has since been revised down by more than a third.
Neither company has left the AI market. What both have done is decide that the AI market is now a managed cost center rather than an open tab.
Why the experiment ended
The economic logic is straightforward at Microsoft and more complicated at Meta.
Microsoft's situation reduces cleanly to portfolio rationalization. The company owns GitHub Copilot, which runs largely on OpenAI models and covers substantially the same coding use case as Claude Code. Paying Anthropic for a product it can sell a version of internally is a defensible experiment and an odd permanent state. When Microsoft's legal team also flagged Anthropic's data retention terms for review, the case for letting the tab run was gone.
Meta's situation carries a second dimension. Claude Code reached the scale it did partly because engineers treated usage as a status signal. An internal leaderboard employees called "Claudeonomics" tracked model consumption; some reportedly engaged in what they termed tokenmaxxing. The behavior is a recognizable product of any system that offers a free resource to engineers: usage expands to fill available budget, and the available budget turns out to be very large when the incentive to constrain it sits nowhere near the person using it.
The distillation question made the business case for intervention even clearer. Meta is building its own coding model, MetaCode, and AI tools developed by a competitor. An internal memo warned that outputs from external models entering training pipelines could constitute model distillation, a concern that became more credible after Anthropic sued Alibaba on related grounds earlier this year. The legal exposure gave the cost argument a risk-management frame that engineering managers could not easily dismiss.
What the numbers say about the broader market
Anthropic's reported revenue tells a different story from the two departures. The company's annualized run rate reached $65 billion in July, up from $47 billion in May and roughly sevenfold from where it ended 2025. The enterprise customers spending through Microsoft's platforms continue to grow. What Meta and Microsoft are cutting is what they pay to give their own employees access — not what their business customers pay to use the models.
The pattern fits a wider shift in how enterprise software budgets are being managed this year. Zylo's 2026 SaaS Management Index found that AI-native application spending jumped 108 per cent year-over-year, reaching an average of $1.2 million per organization and climbing to sixteen per cent of the top fifty enterprise software budgets. The same survey found that seventy-eight per cent of IT leaders reported unexpected charges from consumption-based AI pricing, and sixty-one per cent cut planned projects mid-year as a result.
The gap between what companies budgeted and what AI tools actually cost at scale is the constraint now. A tool that costs $30 per user per month looks manageable at a hundred employees. At sixty thousand, it does not.
The governance layer that was always coming
Zylo's data also found that forty-one per cent of enterprises are actively reducing their application footprint, and that sixty-six per cent now prefer AI integrated into existing platforms over standalone tools. ServiceNow moved in April to make AI capabilities default across its entire product line rather than a separate purchase. Microsoft's own Copilot strategy has embedded AI into Microsoft 365 across every tier.
The logic is not that standalone AI tools are poor. It is that a tool whose value proposition depends on sitting beside an existing workflow is in a weaker position than a tool that becomes the workflow. Engineers prefer the better tool. Finance departments prefer the tool on the existing contract. At the scale where Meta and Microsoft operate, finance wins eventually.
The earlier phase of AI infrastructure spending treated the question of which model to use as secondary to the question of building capacity to use any model. That phase accelerated capital expenditure at a rate few predicted. What is emerging now is the accountability layer: the question of which tools justify their line items when someone is actually reading the bill.
Vendor concentration has been the accompanying risk throughout. The organizations that built deep dependencies on a single external model are now the ones navigating the transition cost. The organizations that treated model procurement as a governance question from the start are finding the current moment less disruptive.
The AI tool budget has found its finance department. That is not a retreat. It is the normal second act of any technology that actually works.
Topics aienterprisetechnologyanthropicmicrosoftmeta
