Why this release matters
On 17 November 2026, at Ignite, Microsoft AI released two coding models of its own: MAI Code 2 and MAI Code 2 Flash. They went live in GitHub Copilot the same day. For a company that has built Copilot on OpenAI and Anthropic models, this is a big step. InfoWorld reads it as part of Microsoft's push to rely less on those two labs. Pareekh Jain, CEO of EIIRTrend & Pareekh Consulting, told InfoWorld that Microsoft gains because it no longer pays an outside lab every time a developer uses Copilot.
Builders mainly want to know what the models cost, where they're strong and where they aren't. We'll start with the price, because that's how Microsoft is selling them.
The pricing, in tokens and credits
GitHub's Copilot pricing reference gives both models a promotional rate until 17 May 2027. After that date the list price applies, and every rate doubles.
- MAI Code 2, promo: US$0.50 per million input tokens, $0.025 per million cached input tokens, $2.50 per million output tokens.
- MAI Code 2, list price from May 2027: $1.00, $0.05 and $5.00.
- MAI Code 2 Flash, promo: $0.05 input, $0.005 cached, $0.25 output per million tokens.
- MAI Code 2 Flash, list price: $0.10, $0.01 and $0.50.
Copilot bills these through AI credits, and one credit equals one US cent. So a million input tokens on MAI Code 2 costs 50 credits at the promo rate. Individual plans come with a set allowance of credits. Business and Enterprise plans pool each seat's allowance at the billing-entity level. Anything over the allowance is billed at the per-token rates above.
Here's a rough example. Say an agent session reads one million tokens of code and writes 100,000 tokens. On MAI Code 2 that's about $0.75 (75 credits) today and $1.50 once the promo ends. On Flash the same session costs about $0.075 today and $0.15 later. If you don't want surprise bills, GitHub suggests setting a Bundled AI credits budget of $0 and turning on "Stop usage when budget limit is reached".
Where you can use them
GitHub's changelog lists MAI Code 2 on Copilot Pro, Pro+, Max, Business and Enterprise. Flash covers those plans and also Copilot Free and Student. Both show up in the model picker in VS Code, Visual Studio, Copilot CLI, the Copilot cloud agent, the GitHub Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode and Eclipse.
Admins should check one thing. Under GitHub's default model enablement, new models switch on automatically unless an admin has turned off the global default or blocked the model. If your organisation vets each model before staff can use it, look at the Copilot model policy now.
Outside Copilot, the options are limited. Flash is in a limited-access preview on Microsoft Foundry. Microsoft says OpenRouter, Vercel, Hermes, OpenCode and Kilo are "coming soon" but gives no dates. At Ignite, Microsoft also announced MAI Licensed Weights, which lets enterprises customise an MAI model and run it in their own Azure tenancy. The Ignite post doesn't say whether the Code models are part of that programme.
What each model is for
Microsoft splits the work between the two models. MAI Code 2 is meant for deep debugging, changes that span many files, reasoning across a whole repository, and front-end work like dashboards and visualisations. Flash is meant for high-volume, lower-stakes jobs: explaining code, small targeted edits, routine fixes and quick back-and-forth iteration.
Microsoft also says what neither model is ready for yet. It hasn't tuned them for fundamental science research or very long-horizon agentic tasks. If you're planning agents that run unattended for hours, keep that in mind.
How the benchmark claims stack up
Microsoft's headline figure averages four tests: DeepSWE 1.1, Terminal-Bench 2.1, SWE-Atlas and FrontierCode v1.1. On that average, it says MAI Code 2 gets 98% of GPT-6.1 Sol's performance for 40% of the cost, and 96% of Claude Opus 5.5's performance for 5% of the cost. InfoWorld published the full table:
| Benchmark | MAI Code 2 | GPT-6.1 Sol | Opus 5.5 | |---|---|---|---| | DeepSWE 1.1 | 73.2% | 74.6% | 69.0% | | Terminal-Bench 2.1 | 84.6% | 85.0% | 78.7% | | SWE-Atlas | 43.1% | 47.4% | 57.5% | | FrontierCode v1.1 | 52.0% | 49.4% | 56.0% | | Average | 63.2% | 64.1% | 65.3% |
The averages are close. SWE-Atlas is the outlier, where Opus 5.5 leads MAI Code 2 by 14.4 points. If your work looks like that benchmark, the headline ratio will flatter Microsoft's model.
Flash looks stronger against its own peers. Across three benchmarks (FrontierCode wasn't included), it averaged 61.0%, against 55.1% for GPT-6 Luna and 53.9% for Claude Haiku 5.5. It led on all three, and its 44.8% on SWE-Atlas is a little higher than the bigger MAI Code 2 scored. Average cost per task tells the same story: Flash $0.052, Luna $0.080, Haiku 5.5 $0.319. In the larger tier the figures were MAI Code 2 $0.309, Sol $0.761 and Opus 5.5 $6.306.
For front-end work, Microsoft ran blind side-by-side tests with 258 annotators. Developers picked MAI Code 2's output over GPT-6 Sol's 64% of the time.
The conditions behind the numbers
Keep the test setup in mind when you read these numbers. Every model ran at maximum reasoning effort. Microsoft ran three of the four benchmarks itself, on its internal MSBench platform through Copilot CLI and without subagents. Only the FrontierCode scores and costs came from Artificial Analysis. The MAI cost figures also use the 50% promo price. Sanchit Vir Gogia, chief analyst at Greyhound Research, told InfoWorld the gap "narrows sharply once the promotion ends."
Jain made a wider point about total cost. If model spend is 10% of what a team pays to build software, a 60% cut in model prices saves only about 6% overall. Integration, testing, security and human review still cost what they did. InfoWorld also says Microsoft's model cards give both models a high rating for offensive cyber capability. Microsoft publishes a model card and data cards on the MAI Code 2 page, and teams should read them before rolling the models out widely.
What builders can do now
- Start with Flash for routine work. It's the cheapest option, and its lead over Luna and Haiku 5.5 is the strongest result in the launch.
- Run MAI Code 2 against your current model on real tickets. The SWE-Atlas gap is a good reason to test first. Gogia suggests tracking cost per merged change, change failure rate and security findings against a baseline.
- Budget at list price. Plan for rates doubling on 17 May 2027 and set Copilot budget caps.
- Check admin settings. Make sure auto-enablement and data settings match your policies.
- Watch for custom versions. Microsoft says AT&T built a custom Flash model for 5G network analysis, and Woven by Toyota fine-tuned Flash on its MISRA fix history. These are partner-reported results, and both are separate from the models in Copilot.
Sources
- Introducing MAI Code 2 and MAI Code 2 Flash (Microsoft AI)
- Models and pricing for GitHub Copilot (GitHub Docs)
- MAI Code 2 and MAI Code 2 Flash in GitHub Copilot (GitHub Changelog, 17 November 2026)
- Microsoft AI at Ignite 2026 (Microsoft AI)
- MAI Code 2 model page (Microsoft AI)
- Microsoft launches MAI Code 2 models to challenge OpenAI and Anthropic in AI coding (InfoWorld)
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.