Microsoft's AI team (MAI) unveiled two coding models at Ignite 2026, MAI Code 2 and MAI Code 2 Flash. Both are already available in GitHub Copilot, and the Flash version is also in limited-access preview on Microsoft Foundry. Microsoft says platforms including OpenRouter, Vercel, OpenCode and Kilo will add support “soon.”
Two models, two jobs
MAI Code 2 is aimed at the heavier end of everyday enterprise coding: multi-file implementation, deep debugging, repository-level reasoning, UI code, and getting code into production. The Flash version handles lighter, high-throughput work such as explaining code, targeted edits, routine fixes and fast iteration. Microsoft is explicit that neither model has been tuned for fundamental scientific research or very long-horizon agentic tasks.
The comparison figures Microsoft provided:
| Comparison | Performance | Cost |
|---|---|---|
| MAI Code 2 vs. GPT 6.1 Sol | 98% | 40% |
| MAI Code 2 vs. Claude Opus 5.5 | 96% | 5% |
| Code 2 Flash vs. GPT-6 Luna | Comparable | 65% |
| Code 2 Flash vs. Claude Haiku 5.5 | Comparable | 20% |
The basis for these numbers needs to be unpacked. The two MAI Code 2 rows are averages across four benchmarks: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, with every model set to its maximum reasoning effort. The first three were run on MSBench, Microsoft's internal evaluation platform, through GitHub Copilot CLI without sub-agents; FrontierCode's scores and costs come from Artificial Analysis. Costs for both MAI models are calculated at a 50% promotional price. The launch page does not say how long the promotion will last, and neither the list price per million tokens nor any individual benchmark score has been published.
A second set of figures comes from a human blind test: 258 annotators compared outputs side by side on front-end visual coding tasks and preferred MAI Code 2 64% of the time against GPT-6 Sol. The tasks covered interfaces, games, data visualization, 3D scenes and SVG graphics.
Same foundation, extended into Office
These models are not only for writing code. Microsoft says MAI Work Flash, optimized from Code 2 Flash, is already used in Microsoft 365 and Fabric IQ: in Word it performs close to GPT 6.1 Sol while using about 42% fewer input tokens per call, and in Teams, instruction following and tool calling save up to 28% of tokens.
Also announced at the event were the image-editing model MAI Image 2.7, the security model MAI Cyber 2 Flash, and MAI Licensed Weights. The latter lets enterprises customize MAI model weights inside their own Azure environments, with the data and fine-tuned results belonging to the customer. Pricing has not been announced and must be negotiated through Microsoft's account teams. One line from the keynote sums up the idea:
“we know you want models that feel truly yours, where you control the weights.”
Connecting the MAI line
Looking back, Microsoft's in-house model effort has moved quickly this year. We previously reported that Microsoft launched seven in-house models at once, with its coding model outperforming Claude Haiku; it then finished training the 2.8T-parameter base model MAI-Base-2 and began shifting AI features in Office toward in-house models. From the stated aim of “reducing dependence on OpenAI” to putting a coding model straight into GitHub Copilot's model list, MAI models now sit in the same dropdown as those from OpenAI and Anthropic, open to direct comparison by users.
Copilot users can switch models themselves, and whether developers actually move over from Claude or GPT will say more than any benchmark table. Microsoft has not disclosed how much of Copilot's usage MAI Code 2 accounts for, nor the regular price after the promotion ends.
Sources: Microsoft AI official blog, Microsoft Ignite 2026 keynote transcript, CocoLoop; the MAI Code 2 launch page footnotes were checked for the four-benchmark combination, the maximum reasoning effort setting and the 50% promotional pricing basis.