Meta and Microsoft Cut Internal Use of Claude to Favor In-House AI

Meta Platforms and Microsoft are working to cut their employees' use of Anthropic's Claude, TechRadar reported on Oct. 6. Both companies are among Anthropic's largest corporate customers, according to The Information.

Meta cut its Claude Code seats, meaning paid licenses, from 60,000 earlier this year to 30,000, TechRadar reported. The company is encouraging staff to use MetaCode, its own coding assistant, instead.

Microsoft reduced its internal spending on Claude by a third, The Information reported. It has asked staff to use Copilot rather than Claude. It is also limiting use of Anthropic's Claude Fable 5 because of Anthropic's new data retention rules, according to Reuters.

The use being replaced is everyday knowledge work. Meta staff use models like Claude for writing code, doing research, and drafting internal and external messages, Business Insider reported. Meta has also given staff access to Google Gemini and OpenAI's GPT-5 alongside its own tools to raise productivity, Business Insider reported. In practice, license counts and default settings decide which model handles code suggestions inside the programming editor, research help, and document drafts.

The practice of using a rival's model is not limited to those two. Some Google DeepMind staff have been allowed to use Anthropic's Claude Code in recent months, Business Insider reported.

The broader context here is useful for teams that run internal AI systems. A large internal rollout is at once a cost, a testing ground, and a way to enforce policy. Shifting tens of thousands of licenses from an outside vendor to internal systems changes the cost of running the models and brings prompts, answers, storage rules, and audit logs under one company's control.

In my view, that explains why seats and spending are the first levers mentioned here. They can be measured and reversed quickly, without waiting for a long comparison of test scores.

Looking at what this means for business adoption, using fewer approved assistants has practical advantages. It simplifies setup in the programming editor, support, login management, and data handling. It also gathers usage data in one place, where platform teams can adjust search, memory limits, and agent behavior. The tradeoff affects developers and researchers who had come to prefer Claude for certain coding or writing tasks. They now have fewer approved options for that preference.

Worth flagging is what this volume of use does to the in-house tools. Heavy daily use reveals slow responses, memory limits, and rough integrations that lab tests miss. If MetaCode and Copilot take over sessions once sent to Claude, that feedback should improve suggestion quality, reliability, and fit with daily work over time. We have seen this pattern before with client-server, mobile, and cloud shifts. The tools staff are required to use tend to improve fastest, and those gains often carry over to outside products built on the same base.