Why Microsoft Is Dropping the Superior Claude Code for Its Own Copilot CLI

License revocation

On May 14, The Verge reported that Microsoft will revoke internal licenses for Claude Code, requiring thousands of developers to switch to GitHub Copilot CLI by the fiscal‑year end on June 30.

Benchmark performance

Claude Code achieves an 80.8 % score on the SWE‑bench benchmark, while the GPT‑4o‑based GitHub Copilot scores 72.5 %, a gap of 8.3 percentage points. For tasks that modify more than five files, Claude Code’s success rate is 23 % higher.

Internal adoption timeline

Claude Code was opened to Microsoft engineers in December 2023. Within six months its user base grew sharply, eroding usage of the internally built GitHub Copilot CLI.

Strategic memo

In an internal memo, Rajesh Jha, head of Experience & Devices, acknowledged Claude Code as an important learning tool but emphasized that Copilot CLI is purpose‑built for Microsoft’s code‑base, security requirements, and engineering standards.

Anthropic partnership context

Microsoft’s partnership with Anthropic includes integration of Claude Sonnet 4.5, Opus 4.1 and Haiku 4.5 into Azure Foundry (a $5 billion project) and Claude‑powered features in Microsoft 365 Copilot. The discontinuation of Claude Code does not affect these collaborations.

Pressure on the GitHub team

Jha’s memo states that making Copilot CLI the best “agentic” programming tool is a joint responsibility of GitHub and Experience & Devices leadership. The GitHub team has delivered “significant improvements” based on internal feedback.

Technical gap

Claude Code supports a million‑token context window, handling roughly 3 000 files in a single session. Copilot CLI’s context limit is 128 K tokens. In cross‑file refactoring or debugging scenarios, Claude Code’s completion rate is 89 % versus 60 % for Copilot CLI.

Acquisition consideration

An anonymous Microsoft insider said the company once evaluated acquiring Cursor to close Copilot’s gaps, but abandoned the plan—potentially to avoid antitrust scrutiny—and instead examined other AI startups.

Usage statistics

Internal data show 91 % of engineering teams use Copilot. Among developers who have tried both tools, 61 % consider Claude Code more accurate for complex debugging and refactoring, while 73 % find Copilot faster for routine code completion.

Tool characteristics

Copilot runs inside the editor with a 100‑300 ms response time, excelling at quick completions. Claude Code runs in the terminal, can operate for hours, and is capable of architectural migrations across dozens of files.

Migration window

The forced migration deadline is less than six weeks. Engineers accustomed to Claude Code must adapt to Copilot CLI, which is still catching up in capability.

Conclusion

Microsoft will continue its cloud partnership with Anthropic, but on the product side it mandates the switch from Claude Code to Copilot CLI, prioritizing internal tooling despite Claude Code’s superior performance in several scenarios.

The Verge report – https://www.theverge.com/tech/930447/microsoft-claude-code-discontinued-notepad

Anthropic official news – https://www.anthropic.com/news/claude-in-microsoft-foundry

SWE‑bench leaderboard – https://www.swebench.com

Code example

来源丨
经授权转自
OSC开源社区（ID：oschina2013）
作者丨大东