Microsoft AI Microsoft Built Its Own Coding Model for GitHub Copilot
MAI-Code-1-Flash is Microsoft's first coding AI trained entirely in-house, without OpenAI's data or technology. It's rolling out to all Copilot plans now, and its benchmark numbers suggest it was built for token efficiency rather than headline scores.
Microsoft announced MAI-Code-1-Flash at Build 2026 on June 2 and has been rolling it out to GitHub Copilot users in VS Code since. It’s the company’s first coding model trained from scratch in-house, without distillation or data from OpenAI.
The architecture is a sparse Mixture-of-Experts model with 137 billion total parameters and a 256,000-token context window. Microsoft says the model was trained directly using GitHub Copilot’s production harnesses, so it learned to work with the surrounding tools and IDE integrations rather than from isolated benchmark prompts.
What the Benchmarks Say
Microsoft’s headline number is 51.2% on SWE-Bench Pro, which is 16 points above Claude Haiku 4.5 (35.2%). It also claims MAI-Code-1-Flash solves harder problems using up to 60% fewer tokens than comparable models on SWE-Bench Verified.
The token efficiency angle matters more than it might seem. GitHub Copilot moved to metered, credit-based billing on June 1. A model that burns fewer tokens per task directly translates to lower credit consumption for users. Microsoft’s framing of the model as “built for production, not benchmarks” fits neatly with that context.
The model uses adaptive reasoning, adjusting its depth based on task complexity. A code completion doesn’t need the same reasoning budget as a full agent run, and the model is supposed to calibrate accordingly.
Rollout
MAI-Code-1-Flash is rolling out to Copilot Free, Student, Pro, Pro+, and Max plans. It shows up in the model picker in VS Code. Microsoft is doing a gradual rollout starting with a limited group before expanding.
It’s positioned as a lightweight option: faster and cheaper to run than the top-tier models, but optimized specifically for Copilot’s workflows. The auto picker can choose it when it fits the task.
Full details are in the Microsoft AI announcement and the GitHub Changelog.