Assess
Microsoft's first in-house coding model, announced at Build 2026 on June 2, 2026. MAI-Code-1-Flash is a sparse MoE model (137B total parameters, 256K context) that outperforms Claude Haiku 4.5 by 16 points on SWE-bench Pro and solves tasks with up to 60% fewer tokens — at a mid-tier price of $0.75/$4.50 per million input/output tokens.
Why It's in Assess
MAI-Code-1-Flash marks Microsoft's first proprietary coding model trained from scratch on enterprise-licensed data, with no distillation from any third-party model. After years of reselling OpenAI and Anthropic models through Azure and Copilot, Microsoft now controls its own model weights for the Copilot coding path:
- SWE-bench Pro 51.2% — 16 points above Claude Haiku 4.5 (35.2%) at a comparable price point; strong for a mid-tier model
- Token efficiency: Solves coding tasks with up to 60% fewer tokens on SWE-bench Verified compared to similarly-priced models — a meaningful cost multiplier at API scale
- 256,000-token context window — longer than most mid-tier models; handles large repository context without chunking
- Copilot-native launch: Rolled out June 2 to GitHub Copilot Free, Pro, Pro+, and Max plans in VS Code — the first Microsoft-built model available to Copilot subscribers as an alternative to OpenAI/Anthropic options
It sits in Assess rather than Trial because:
- Direct API access outside Copilot is limited to 3 announced distribution partners as of June 2026; broader Azure AI Foundry availability is still pending
- Pricing described as "still being finalized" per the model card
- Benchmark comparisons are Microsoft-published; independent third-party evaluation (LiveCodeBench, MBPP, real-world coding scenarios) is pending
Key Characteristics
| Property | Value |
|---|---|
| Architecture | Sparse MoE, 137B total parameters |
| Context window | 256,000 tokens |
| Input pricing | $0.75 / million tokens |
| Cached input | $0.075 / million tokens |
| Output pricing | $4.50 / million tokens |
| SWE-bench Pro | 51.2% |
| Provider | Microsoft |
| Availability | GitHub Copilot (VS Code, June 2, 2026); 3 distribution partners |
| Training data | Enterprise-licensed; no third-party model distillation |
What to Watch
- Direct API availability: MAI-Code-1-Flash is primarily Copilot-native as of launch; broader Azure AI Foundry API access is not yet live. Watch for an Azure OpenAI or AI Foundry release announcement.
- MAI-Thinking-1 companion: Announced alongside Flash, MAI-Thinking-1 is a 35B-active-parameter reasoning model in private preview on Microsoft Foundry with a 256K context window (multiple secondary sources; primary microsoft.ai announcement blocked). Public preview on MAI Playground expected soon (as of June 2026). If both prove competitive, Microsoft's reliance on OpenAI/Anthropic for Copilot capability will fall substantially.
- Independent benchmarks: SWE-bench Pro 51.2% comes from Microsoft's own evaluation. Third-party benchmarking on standard open leaderboards will determine whether these numbers hold outside Microsoft's test harness.
- Pricing finalization: The model card explicitly notes pricing is not yet final — worth rechecking before committing to cost-based capacity planning.