At a glance: MAI-Code-1-Flash, announced June 2, 2026 at Microsoft Build. First Microsoft-trained model built using GitHub Copilot’s production tool harnesses. 85.8% on Microsoft adversarial coding benchmark, ~51% SWE-Bench Pro. 60% fewer tokens on complex tasks vs. comparable models. Live now in the Copilot model picker — Free, Pro, Pro+, and Max tiers. Third-party access via Fireworks AI, Baseten, OpenRouter. Part of our Builder’s Log.

Microsoft has spent two years watching Cursor and Claude Code absorb the developer mindshare that GitHub Copilot should have owned by default. Copilot invented the AI pair-programmer category in 2021, had the VS Code integration, had GitHub’s codebase signal — and still lost the early-adopter developer to whoever had the best model that week.

MAI-Code-1-Flash is Microsoft’s answer. Not just a new model for Copilot to use, but a model trained inside Copilot’s production harness — which is a different thing and the part worth paying attention to.

What Makes This Different From Other Coding Models

Most coding model benchmarks are evaluated against coding tasks: solve a GitHub issue (SWE-bench), pass unit tests, implement a function from a docstring. The model is trained, then evaluated on those tasks, then deployed to tools like Copilot.

MAI-Code-1-Flash inverts that. Microsoft trained it with the actual GitHub Copilot production tool harnesses — the multi-step file editing, terminal calls, context retrieval, and inline chat flows that real Copilot users run daily. The model learned to be good at Copilot workflows, not just good at coding benchmarks that approximate Copilot workflows.

The practical consequence Microsoft claims: up to 60% fewer tokens consumed on complex coding tasks versus comparable models. In a billing environment where Copilot switched to usage-based AI Credits on June 1, token efficiency is not an abstract engineering achievement. It is a dollar amount.

The Benchmark Picture

Microsoft has published results on two benchmarks:

Microsoft adversarial coding benchmark (internal)

  • 186 questions across 34 categories
  • Designed to include inverted problems, impossible tasks, and underdetermined scenarios — not just standard coding exercises
  • MAI-Code-1-Flash: 85.8% adjusted accuracy
  • Described as outperforming Claude Haiku 4.5

SWE-bench Verified

  • Microsoft’s official claim: “at or above Claude 3.7 Sonnet”
  • Community-surfaced number from the model card: ~51% on SWE-Bench Pro

Interpret these carefully. The 85.8% figure is on an internal Microsoft benchmark that has not been independently reproduced. The SWE-Bench Pro 51% is more comparable — it puts MAI-Code-1-Flash in the same tier as GPT-5.3, but below Kimi K2.6 (~58.6%), GLM-5.1 (~58.4%), and Claude Opus 4.6’s baseline performance.

For context on the rest of the MAI family: MAI-Thinking-1 (the reasoning model) is described as “competitive with Claude Opus 4.6 on SWE-Bench Pro.” MAI-Code-1-Flash is not claiming that level — it is positioned as an efficient everyday-coding model, not a frontier-reasoner.

The honest read: MAI-Code-1-Flash is strong in its efficiency tier. It is not trying to claim the absolute top of the leaderboard, which is currently occupied by Kimi K2.6, GLM-5.1, and the Anthropic Opus line. It is trying to be the right model for the volume of coding work that flows through Copilot every day.

Architecture: What Microsoft Has and Has Not Disclosed

Microsoft has confirmed:

  • Adaptive solution length control — the model adjusts how much reasoning budget it spends based on task complexity. A simple variable rename uses minimal compute; a multi-file refactor gets more.
  • Commercially licensed training data — same claim as MAI-Thinking-1; built to avoid the IP provenance issues that plagued DeepSeek R1.
  • Production harness training — trained with real Copilot tool interactions, not a simulation.

Microsoft has not disclosed:

  • Parameter count
  • Architecture type (dense vs. MoE)
  • Training infrastructure

The “Flash” naming convention mirrors what Anthropic uses for smaller/faster Claude variants and what Google uses for the Gemini Flash tier. Microsoft appears to be establishing that MAI-Code-1-Flash is the speed/efficiency variant of a larger MAI-Code-1 family — but no larger variant has been announced yet.

Access: Where You Can Actually Use This

GitHub Copilot (live now) MAI-Code-1-Flash is in the Copilot model picker in VS Code as of June 2, 2026. The rollout started with a limited percentage of users and is expanding. Available on all paid Copilot tiers: Free, Pro ($10/month), Pro+ ($39/month), and Max. On paid tiers it can also be selected by the automatic model picker when Copilot decides it’s the right tool for the current task.

Under the usage-based billing model that started June 1, MAI-Code-1-Flash’s token efficiency advantage translates directly to credit preservation. If the 60% token reduction claim holds at production, agentic Copilot workflows that would otherwise exhaust a Pro plan’s $39 monthly credit budget in two days might stretch to five.

Third-party API access MAI-Code-1-Flash is available outside of Copilot through three distribution partners announced at Build 2026:

  • Fireworks AI — inference API, per-token pricing
  • Baseten — model serving platform
  • OpenRouter — unified model router

Azure AI Foundry pricing for the broader MAI model family exists but specific per-token rates for MAI-Code-1-Flash were not published at launch.

GitHub Models GitHub’s free model sandbox provides prototyping access without an Azure subscription. Available now with a GitHub account.

Completing the MAI Stack

Build 2026 launched three distinct MAI model lines in the same week:

Model Category Primary access Notable claim
MAI-Code-1-Flash Coding GitHub Copilot model picker 60% token reduction vs. comparable models
MAI-Thinking-1 Reasoning Azure AI Foundry, GitHub Models Not distilled from any other model
MAI-Image-2.5 / MAI-Voice-2 / MAI-Transcribe-1.5 Multimodal Azure AI Foundry, Office apps Hear / see / speak stack unified

The pattern is visible: Microsoft is building a parallel model stack that does not depend on OpenAI for any capability. In April 2026, restrictions in the Microsoft-OpenAI partnership were lifted, giving Microsoft the right to serve its own models in products rather than defaulting to OpenAI. Build 2026 is the first full public exercise of that right.

What Builders Should Actually Watch

The token efficiency claim is verifiable. If you are a heavy Copilot user, you can select MAI-Code-1-Flash from the model picker today and run it against your real workflows. The claim is 60% fewer tokens on hard tasks — that’s measurable in your usage dashboard under the new AI Credits billing. This is a more actionable signal than benchmark numbers.

The SWE-Bench Pro 51% number anchors expectations. This is a good everyday coding model, not a state-of-the-art agentic coder. For fully autonomous issue resolution on complex open-source repositories, Kimi K2.6, GLM-5.1, or Claude Opus 4.6 remain stronger. MAI-Code-1-Flash is the right choice for the inline chat, autocomplete, and quick-fix workflows that represent the majority of Copilot usage.

The training environment matters for Copilot specifically. A model trained with Copilot’s actual tool harnesses should behave more reliably inside Copilot’s agentic workspace than a model benchmarked externally and then deployed into it. Whether that holds in practice is something only usage will surface.

The third-party distribution question is open. Fireworks, Baseten, and OpenRouter give access outside of GitHub, but pricing and rate limits on those channels aren’t published. If you need MAI-Code-1-Flash for a product you are building (not just for your own development workflow), watch those platforms for pricing announcements.