Microsoft AI has launched MAI Code 2 and MAI Code 2 Flash for enterprise coding.
Both models are available now in GitHub Copilot. Microsoft also plans to license MAI model weights, letting companies customise and run their own versions in Azure, but that programme isn’t generally available yet.
Microsoft AI unveiled the coding models at Ignite in San Francisco on Tuesday, 17 November. The coding models were part of a wider MAI launch: WindowsForum counts seven items, including product integrations, while Axios counts six new models.



Microsoft’s cost comparisons use GPT-6.1 Sol and GPT-6 Luna, even though OpenAI’s GPT-6.2 Astra and Sol reached ChatGPT and GitHub Copilot the same day.
What Microsoft’s coding benchmarks show
Across four benchmarks, Microsoft says MAI Code 2 delivered 98% of GPT-6.1 Sol’s performance at 40% of the cost and 96% of Claude Opus 5.5’s performance at 5%. The tests were DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1.
In three of the company’s coding benchmarks, Flash performed comparably to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their respective costs. All models ran at maximum reasoning effort, and the MAI Code 2 cost figures reflect a 50% promotional price. The cost comparisons depend on list prices and token mixes Microsoft hasn’t fully detailed, and no independent validation has been published yet.
Unite.AI broke down Microsoft’s figures by average pass rate and cost per task:
| Comparison | Model | Average pass rate | Average cost per task (US dollars) |
|---|---|---|---|
| Code 2 | MAI Code 2 | 63.2% | $0.309 |
| Code 2 | GPT-6.1 Sol | 64.1% | $0.761 |
| Code 2 | Claude Opus 5.5 | 65.3% | $6.306 |
| Flash | MAI Code 2 Flash | 61.0% | $0.052 |
| Flash | GPT-6 Luna | 55.1% | $0.080 |
| Flash | Claude Haiku 5.5 | 53.9% | $0.319 |
In a separate blind frontend evaluation with 258 annotators, reviewers preferred MAI Code 2’s output to GPT-6 Sol’s 67% of the time. Microsoft says neither coding model is tuned for fundamental science research or for AI agents to work through very long tasks.
Where the MAI models are available
MAI Code 2 is available on GitHub Copilot Pro, Pro+, Max, Business, and Enterprise. Flash is also available on Free and Student. Both use usage-based billing and are available to eligible annual subscribers. Microsoft hasn’t confirmed UAE availability for either model.
Business and Enterprise admins can control access in Copilot’s model policy. New models are enabled automatically unless an admin turns off the global default or disables the model. Flash is also in limited-access preview in Microsoft Foundry and will reach OpenRouter, Vercel, Hermes, OpenCode, and Kilo.
Microsoft adapted Flash into MAI Work Flash for Microsoft 365 and Fabric IQ. In the company’s tests, it matched GPT-6 Luna in Excel and PowerPoint and performed similarly to GPT-6.1 Sol in Word with about 42% fewer input tokens per call. On Teams tasks, it beat Luna with up to 28% fewer tokens. In Fabric IQ, it outperformed GPT-5.6-sol, GPT-6-astra, and Opus-5.5 at generating DAX queries.
Microsoft will license weights for use in Azure
MAI Licensed Weights will let organisations adapt Microsoft’s models to their own data and deploy them in their own Azure environments. The Information reports Microsoft will license the weights, the settings that determine how a model answers, so customers can build customised versions. The report says neither Anthropic nor OpenAI lets customers customise models this way.
The programme is coming soon through Frontier Company engagements, and interested customers can contact their Microsoft account team. Microsoft says customers retain ownership of their data and adaptations, and it will only use them to train its own models with express permission.
In Microsoft’s customer evaluations, MAI outperformed GPT-6.1 Sol on AT&T’s standard 5G network analytics tasks at one-fiftieth of the cost. Woven by Toyota got fixes ready to merge in 44% of cases where frontier models needed human intervention. Alaska Airlines improved the accuracy of aircraft maintenance report classification from 81% to 91%.
Following its earlier MAI voice models, Microsoft now has its own coding models alongside those from its AI partners. Mustafa Suleyman predicted significant adoption but declined to set a target for how much Copilot traffic would shift to MAI models.


















