$ cat wiki/models/mai-code-1.md
MAI-Code-1 / MAI-Code-1-Flash
modelupdated 2026-07-21created 2026-06-02
Spec
| Attribute | Value |
|---|---|
| Developer | Microsoft (MAI team, in collaboration with the GitHub Copilot team) |
| Released | 2026-06-02 (MAI-Code-1-Flash; full-size MAI-Code-1 GA not announced) |
| Announced | 2026-06-02 (Microsoft Build 2026 keynote) |
| Context window | unknown |
| Pricing | unknown |
| License | Closed (proprietary) |
| Availability | MAI-Code-1-Flash: immediate GA (deployed to the Copilot model picker on the day of the Build keynote) |
| Architecture | Coding-specialized (trained inside the Copilot production harness) |
| Parameter Class | ~5B (MAI-Code-1-Flash) |
| Distribution Channels | GitHub Copilot (Free/Pro/Pro+/Max model picker), Azure AI Foundry |
Release Date
- MAI-Code-1-Flash: 2026-06-02 (immediately available)
- MAI-Code-1 (full size): GA date not announced
What Is It
A coding-specialized model that Microsoft trained inside GitHub Copilot's actual production harness, optimized directly for Copilot usage patterns.
Key features:
- Efficiency: 60% fewer tokens on hard tasks compared to similar models
- Copilot-native: covers GitHub Copilot inline completion + chat + agent workflows
- Immediate deployment: accessible via the model picker across all Copilot tiers on day one
Benchmarks
| Benchmark | Score |
|---|---|
| Microsoft Adversarial Coding Benchmark | 85.8% |
| SWE-Bench Pro | ~51% |
⚠️ Benchmarks are Microsoft's own measurements. Independent verification not yet complete (as of 2026-06-02).
Comparison note: Claude Opus 4.8 scores 69.2% on SWE-bench Pro (MAI-Code-1-Flash is competitive on a small-model basis)
Use Cases
- GitHub Copilot code completion (inline)
- Copilot chat (VS Code, JetBrains, GitHub.com)
- Autonomous agent tasks within Copilot Workspace
- GitHub Actions Copilot pipelines
Compared To
| Model | SWE-bench Pro | Notes |
|---|---|---|
| Claude Opus 4.8 | 69.2% | Anthropic's strongest, much larger |
| Devstral 2 | 72.2% | Mistral coding-specialized |
| GPT-4.1 (OpenAI) | ~65% | OpenAI's coding baseline |
| MAI-Code-1-Flash | ~51% | small (~5B class), efficiency-focused |
Significance
- Instant distribution channel: GA to tens of millions of GitHub Copilot developers on day one — an immediate adoption pipeline no competitor has
- Copilot-native training: trained directly in the production environment → reflects real usage patterns
- Efficiency: 60% token savings → more tasks handled on the same credit budget
Open Questions
- MAI-Code-1 (full size) parameter count undisclosed
- Awaiting independent benchmark verification
- Relationship to Project Polaris: Polaris (MoE, Aug 2026 GA) is the eventual default model, with Code-1-Flash possibly coexisting as a lightweight, fast alternative
Related
- Microsoft
- Project Polaris — GitHub Copilot's eventual default model (Aug 2026 GA)
- MAI-Thinking-1 — reasoning model announced at the same time
- Agents (LLM Agents) — Copilot Workspace agents
- Software 3.0