← Back to feed

Microsoft shipped MAI-Code-1-Flash on June 2, a coding model built entirely by Microsoft using clean and appropriately licensed data. It is rolling out now to GitHub Copilot individual users in VS Code via the model picker and the default auto picker.

The headline benchmark is SWE-Bench Pro: 51.2% versus Claude Haiku 4.5’s 35.2%, a 16-point gap. The comparison is the relevant one — MAI-Code-1-Flash is positioned as Haiku 4.5’s replacement in the Copilot stack, not a challenger to frontier models. It also outperforms Haiku 4.5 on SWE-Bench Verified, SWE-Bench Multilingual, and Terminal Bench 2, all measured in the same production harness developers use.

Efficiency is the actual story

Microsoft ran a dual metric: pass rate and average tokens per completed task. On that combined view, MAI-Code-1-Flash lands in what its team calls the “Ideal Zone” — higher accuracy than Haiku 4.5 and substantially lower token consumption.

The token reduction is not incidental. The model was trained with adaptive solution length control: it stays concise for simple edits and allocates more reasoning budget to complex, multi-file tasks. On SWE-Bench Verified, MAI-Code-1-Flash solves harder tasks using up to 60% fewer tokens than the baseline. For developers on token-metered plans — which now covers all GitHub Copilot tiers after the June 1 billing change — that efficiency gap translates directly to cost.

Instruction following and agentic tasks

On IF Bench (instruction following), MAI-Code-1-Flash leads Claude Haiku 4.5 by 28.9 points — the widest margin across any benchmark tested. The narrowest is Advanced IF at +14.5. Agentic tool use, measured on tau-1-Bench, also shows improvement.

The model was trained directly on the GitHub Copilot production harness, including repository question answering, refactoring tasks, and telemetry-grounded tasks pulled from real developer workflows. That training alignment between evaluation and production is the methodological bet behind the efficiency claims.

Context

MAI-Code-1-Flash is the second Microsoft in-house model to ship in recent months, following MAI-Image-2.5 in late May. The pattern is a direct extension of what Microsoft announced at Build 2026: breaking OpenAI dependency for specific workloads while keeping OpenAI models available at the premium tier. At the developer assistance layer — the highest-volume Copilot use case — Microsoft now has its own model in the default rotation.

Pricing for MAI-Code-1-Flash was not disclosed in the announcement; the model ships under Copilot’s token-metered billing structure. Given the 60% token reduction claim, effective cost per completed task should be meaningfully lower than Haiku 4.5 even at equivalent per-token rates.

Key numbers

• SWE-Bench Pro: 51.2% (vs. Claude Haiku 4.5: 35.2%, +16 pts)
• IF Bench: +28.9 points over Haiku 4.5
• Token efficiency: up to 60% fewer tokens on SWE-Bench Verified
• Deployment: GitHub Copilot VS Code model picker, rolling out now
• Training: end-to-end Microsoft, clean licensed data