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.
Update (2026-08-22): This model has been superseded and is being retired. Microsoft shipped MAI-Code-1.1-Flash on August 11, 2026, and GitHub has scheduled MAI-Code-1-Flash for retirement across all Copilot experiences on September 10, 2026. Per Microsoft’s own announcement, MAI-Code-1.1-Flash adds native vision/image support, is 25% more token-efficient, improved 22% on Terminal-Bench 2.1 and 15% on .NET tasks, and lists at a 73% lower price than MAI-Code-1-Flash — GitHub’s rate card puts MAI-Code-1.1-Flash at $0.20/M input and $1.20/M output tokens, versus $0.75/$4.50 for MAI-Code-1-Flash. The rest of this article describes the model as it launched at Build 2026 and is kept for the historical record; for current guidance, use MAI-Code-1.1-Flash.
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 to solve harder problems on SWE-Bench Verified 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 benchmark families, both in its own launch post:
Microsoft adversarial coding benchmark (internal)
- 186 questions across 34 categories, built around “adversarial traps like inverted classics, impossible tasks, and underdetermined scenarios”
- MAI-Code-1-Flash: 85.8% adjusted accuracy
- Microsoft’s own wording: “MAI-Code-1-Flash surpasses Claude Haiku 4.5 overall” on this benchmark
SWE-Bench Pro (Microsoft’s own comparison, not community-surfaced)
- Microsoft’s published number: 51.2% for MAI-Code-1-Flash vs. 35.2% for Claude Haiku 4.5 — a +16-point lead, reported directly by Microsoft alongside SWE-Bench Verified, SWE-Bench Multilingual, and Terminal-Bench 2 results (also claimed wins over Claude Haiku 4.5 on all four)
- Microsoft’s post does not compare MAI-Code-1-Flash to Claude 3.7 Sonnet on SWE-Bench Verified or any other benchmark; an earlier version of this article attributed that comparison to Microsoft and it did not check out against the source, so it has been removed
Interpret these carefully. The 85.8% figure is on an internal Microsoft benchmark that has not been independently reproduced. The SWE-Bench Pro number is more comparable to the rest of the field: on the independent SWE-Bench Pro leaderboard, MAI-Code-1-Flash’s 51.2% sits below GPT-5.3 Codex (56.8%), Kimi K2.6 (58.6%), and GLM-5.1 (58.4%) — all measured on the same public leaderboard. Claude Opus 4.6 is not listed on that leaderboard (only the newer Opus 4.7/4.8), and other SWE-Bench Pro variants (different scaffolds and data splits) put Opus 4.6 anywhere from below to above MAI-Code-1-Flash’s score, so that comparison is cut here rather than guessed at.
For context on the rest of the MAI family: Microsoft says MAI-Thinking-1 (the reasoning model) matches 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.
- 5 billion active parameters — disclosed in Microsoft’s Build 2026 model-family recap, which describes MAI-Code-1-Flash as “comparable to Haiku but cheaper.” Microsoft uses “active parameters” phrasing but has not used the words “Mixture-of-Experts” or “dense” for this specific model, and has not disclosed a total parameter count.
- Training data described as “clean, traceable and enterprise-grade,” without distillation from third-party models — Microsoft’s own phrasing for this model; Microsoft has not used the term “commercially licensed” specifically for MAI-Code-1-Flash (it has used similar language for MAI-Thinking-1), so this article no longer repeats that specific term here.
- Production harness training — trained with real Copilot tool interactions, not a simulation.
Microsoft has not disclosed:
- Total parameter count (only the 5B active-parameter figure)
- Confirmation of a Mixture-of-Experts vs. dense architecture, in explicit terms
- 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. Availability started with a limited set of users and is expanding gradually. It is rolling out across Copilot Free, Student, Pro, Pro+, and Max plans — Pro is $10/month and Pro+ is $39/month. 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. A Pro+ plan’s $39/month credit budget (7,000 credits, per GitHub’s published allotments) stretches further if the 60% token reduction holds in production agentic sessions — the exact number of extra sessions depends on your workflow, so treat this as directional rather than a fixed multiplier.
Third-party API access MAI-Code-1-Flash is available outside of Copilot through three distribution partners Microsoft named in its Build 2026 model-family recap: Fireworks AI, Baseten, and OpenRouter.
On the Copilot side, GitHub’s own rate card lists MAI-Code-1-Flash at $0.75 per million input tokens ($0.075 per million cached) and $4.50 per million output tokens — as of this audit (2026-08-22) that rate card still lists the model as generally available, but GitHub has confirmed it retires on September 10, 2026 in favor of MAI-Code-1.1-Flash, which the same rate card prices at $0.20 per million input tokens and $1.20 per million output tokens. Per-token pricing on Fireworks AI, Baseten, and OpenRouter was not independently confirmed as of this audit — check each platform directly before budgeting against it.
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 | Private preview, Microsoft Foundry | 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. On April 27, 2026, Microsoft and OpenAI restructured their partnership, making Microsoft’s license to OpenAI’s IP non-exclusive and freeing Microsoft to ship its own models rather than defaulting to OpenAI’s. Build 2026 is the first full public exercise of that new latitude.
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 and GLM-5.1 score higher on the public SWE-Bench Pro leaderboard. 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.
Related Coverage
- MAI-Thinking-1 — Microsoft’s reasoning model from the same Build 2026 announcement; designed for enterprise architecture and compliance workflows.
- MAI Multimodal Stack — MAI-Image-2.5, MAI-Voice-2, and MAI-Transcribe-1.5 complete the Build 2026 model launches.
- GitHub Copilot AI Credits Billing — The usage-based billing change that took effect June 1 makes MAI-Code-1-Flash’s token efficiency directly relevant to cost.
- Microsoft Is Building Its Own Coding Model — Our pre-Build preview, written May 29, of what to expect at Build 2026.
- Windows AI Models: Aion 1.0 — The on-device AI story from the same Build 2026 week.
Sources: Microsoft AI on MAI-Code-1-Flash; MAI-Code-1-Flash model page; The Robotics Media on the launch.