← Back to feed

Meta and Microsoft have each taken deliberate steps to reduce employee reliance on Anthropic’s Claude during 2026, according to a report from The Information. The reductions are in active usage, not contracts — neither company has terminated its enterprise agreement with Anthropic. At Meta, active internal users of Claude Code reportedly fell to around 30,000, down from roughly 60,000 earlier in the year, according to The Information.

The shift is driven by two pressures pulling in the same direction: cost management and competitive positioning.

The Cost Calculation

At enterprise scale, AI spend compounds quickly. Claude usage billed by token volume across tens of thousands of engineers adds up to a meaningful line item on a quarterly budget. Both Meta and Microsoft have large internal engineering workforces and, critically, they both now have alternative AI tooling they control.

Reducing external vendor dependency is not the same as deciding the external vendor is inferior. It is a normal procurement decision: when internal alternatives reach sufficient capability, the argument for routing internal employees through an external paid API weakens.

What Each Company Is Moving Toward

Microsoft’s internal alternative is its own Copilot stack, which runs on GPT-series models via its Azure arrangement with OpenAI — a relationship Microsoft has a structural interest in prioritising. Directing internal engineering toward Copilot rather than Claude consolidates model usage in a way that reinforces Microsoft’s own platform and vendor positioning.

Meta’s situation is different. The company has invested heavily in its own Llama model family and has internal AI tooling built around it. Routing internal developers away from Claude toward Llama-based tools is both cheaper and strategically aligned with Meta’s open-source model positioning.

What This Is Not

The reporting is careful on one point: these are usage reductions, not contract cancellations. Anthropic still has enterprise agreements with both companies. The practical implication is that Anthropic retains the revenue from those agreements while losing the deployment footprint that would normally accompany them.

This matters for Anthropic because enterprise footprint is not just revenue — it is the data signal, feedback loop, and reference case that feeds into future contract negotiations and pricing power. A customer that is technically under contract but routing employees away from the product is a different kind of customer than one using it at scale.

Broader Pattern

This story fits a pattern that has been visible at other large technology companies in 2026: initial wide deployment of external AI tools across internal teams, followed by a rationalisation phase in which usage concentrates on workflows where proprietary alternatives are weakest. The rationalisation phase is not a repudiation of the external tool; it is a sign that the enterprise AI market is maturing into something more selective.

For Anthropic, the risk is not that Meta and Microsoft are abandoning Claude — it is that the growth story for enterprise seat count has a ceiling that internal AI tooling is actively lowering.