Breaking

Microsoft launches Copilot-tuned coding AI MAI-Code-1-Flash

June 3, 2026 at 12:45 EDT

Microsoft AI on June 2, 2026 unveiled MAI-Code-1-Flash, a small coding model optimized specifically for GitHub Copilot, and began a phased rollout for Visual Studio Code. It targets individual users (Free/Pro/Pro+/Max) and is available via the model picker and the default Auto picker with no extra configuration. The announcement was published simultaneously on the Microsoft AI blog and the GitHub Changelog.

June 2, 2026 · Microsoft AI

MAI-Code-1-Flash: a small model built to out-code its rivals

Microsoft's first dedicated MAI coding model — trained inside the production GitHub Copilot harness — lands in Visual Studio Code, pairing compact size with benchmark scores that top its comparison target while burning far fewer tokens.

active parameters
(137B total, likely MoE)

~60%

fewer tokens used on SWE-Bench Verified

71.6%

SWE-Bench Verified score

SWE-Bench Pro — like-for-like in VS Code

Resolution rate, same harness · MAI-Code-1-Flash vs comparison target

MAI-Code-1-Flash51.2%

Claude Haiku35.2%

Takeaway: a 5B-active model resolves +16 points more issues than its lightweight comparison target — at lower token cost.

Praised

Beats Claude Haiku on benchmarks
Token-efficient & practical for daily Copilot work
Right-sized small model with Adaptive thinking

Caveats

Individual VS Code users only at launch
Enterprise plans & other editors (e.g. Cursor) not yet
Long-term reviews still scarce; benchmark debate ongoing

The strategy

A shift toward a Microsoft-controlled reasoning stack inside Copilot.

Reduce OpenAI dependence → Native VS Code integration → Lower cost, own coding model

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.