Microsoft dropped two coding models at Ignite today and led with an aggressive number: 96% of Claude Opus 5.5’s performance at 5% of the cost. That’s the claim for MAI Code 2, now selectable in GitHub Copilot across all major IDEs. If that claim holds up in independent testing, it reshapes the economics of running coding agents at scale. The lighter sibling — MAI Code 2 Flash — lands on the Free plan, making it accessible to every Copilot user starting today.
Two Models, Two Jobs
Microsoft is shipping this as a split product, and the split is clean. MAI Code 2 (the full model) targets complex, sustained coding work: deep debugging, multi-file implementation, repository-wide refactoring, and frontend UI development where blind evaluators preferred it over GPT-6 Sol in 67% of tests. It has a 1-million-token context window (922K input, 128K output) and is built to handle the kind of tasks where going back and forth costs more than the model itself.
MAI Code 2 Flash is a 12-billion-parameter sparse mixture-of-experts model, distilled from a larger teacher model, and built for volume. Code explanations, targeted edits, routine fixes, rapid iteration — the kind of work that happens dozens of times per hour in an agent loop. It carries the same 1M-token context window, which is unusual for a model at this size and price point.
Plan Breakdown: Flash Is the News for Free Users
MAI Code 2 Flash is available on every GitHub Copilot plan including Free and Student. No other model claiming Haiku-class performance sits at the free tier. That’s the headline for indie developers and students: a fast, capable coding model at no cost. MAI Code 2 (the full model) requires Pro or above.
Both models are live across VS Code, Visual Studio, JetBrains IDEs, Xcode, Eclipse, Copilot CLI, GitHub.com, and GitHub Mobile. Select the model from the Copilot model picker — no migration required.
What the Benchmarks Actually Show
Microsoft’s claims compare favorably against GPT-6.1 Sol and Claude Opus 5.5. Independent data from benchlm.ai tells a more precise story:
| Model | DeepSWE 1.1 | Terminal-Bench 2.1 | SWE-Atlas | Avg Pass Rate | Cost/Task |
|---|---|---|---|---|---|
| MAI Code 2 | 73.2% | 84.6% | 43.1% | 63.2% | $0.309 |
| MAI Code 2 Flash | 59.9% | 78.3% | 44.8% | 61.0% | $0.052 |
The 2.2-point pass rate gap between the full model and Flash is worth noting — at 6x the cost per task. For most teams, Flash will be the right default for routine coding work. The full model earns its keep on complex architectural decisions where accuracy costs more than the token price.
One honest caveat from Microsoft: neither model is tuned for fundamental science research or long-horizon agentic work. Worth flagging if your use case involves extended planning chains.
The Economics of Flash at Scale
Flash’s promotional pricing through May 17, 2027 sits at $0.05 per million input tokens and $0.25 per million output tokens. The $0.052 average cost per task translates directly: a team running 10,000 agent tasks per day pays roughly $520 per day for Flash. Compare that against models an order of magnitude more expensive, and the case for adopting Flash as the default for routine agent loops becomes straightforward.
After May 2027, pricing doubles — $0.10 input and $0.50 output per million tokens. If you’re evaluating whether to build workflows around Flash, the promo window is a real incentive to move now rather than wait for the community to validate it.
The Realistic Take
Microsoft’s self-reported benchmark comparisons always come with a grain of salt — the company picks the benchmarks, the prompts, and the conditions. What’s credible here is the independent cost-per-task data and the fact that Flash is on the Free plan immediately. The performance claims need a few weeks of community stress-testing before you rebuild your stack around them.
Use Flash now. It’s free tier, production-deployed today, and the downside of trying it is zero. Treat the full MAI Code 2 as a production option worth revisiting once independent benchmark coverage catches up with Microsoft’s numbers.

