Microsoft's new MAI Code 2 and Code 2 Flash target everyday enterprise coding at far lower cost than rivals; available now in GitHub Copilot.

Copy, download or open this article in ChatGPT or Claude
Microsoft announced two new coding models at Ignite on Nov 17, 2026: MAI Code 2 and MAI Code 2 Flash. Both target the routine coding work enterprises do every day, and Microsoft's pitch is strong performance at a much lower price than competing models.
MAI Code 2 sits at the top end of the price-performance curve for common enterprise coding tasks. Microsoft says it handles taking code to production, deep debugging, multi-file implementation, repo-wide reasoning and UI work. The company also says it produces changes that code maintainers would actually merge, at a fraction of the cost of frontier models.
MAI Code 2 Flash is the lighter option. It's designed for high-volume jobs like code explanations, targeted edits, routine fixes and rapid iteration, and it comes in at a much lower cost.
There's a clear limit here. Microsoft says the models were built around everyday coding work and are explicitly not yet tuned for harder problems, such as fundamental science research or very long-horizon agentic tasks.
Microsoft averaged results across four widely used benchmarks that reflect the everyday enterprise tasks it sees in GitHub Copilot: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. On that average, MAI Code 2 delivers:
MAI Code 2 is also tuned for dashboards, visualizations and visual coding more broadly. In blind frontend app evaluations, reviewers preferred its output over GPT-6 Sol 64% of the time.
For MAI Code 2 Flash, Microsoft reports performance comparable to GPT-6 Luna and Claude Haiku 5.5 across three benchmarks covering routine, high-volume coding. The cost is 65% of GPT-6 Luna's and 20% of Claude Haiku 5.5's.
One caveat from Microsoft's own notes: these results come from specific coding tasks and configurations. Performance and cost comparisons shift depending on workload, model configuration, reasoning effort and pricing assumptions.
Microsoft is also using the technology internally. It took Code 2 Flash and further optimized a single model, MAI Work Flash, to serve multiple products across M365 and Fabric IQ. Here's how the company says it performs:
Both MAI Code 2 and Code 2 Flash are available in GitHub Copilot now. Code 2 Flash is also in limited-access preview in Microsoft Foundry.
Microsoft says the models will soon come to Open Router, Vercel, Hermes, Open Code and Kilo. Teams interested in customizing MAI Code 2 can find details on the company's Custom Models page.
What stands out to me here is not so much the image model itself, but the speed. The MAI family only started at Build in June, with seven models for reasoning, code, images, speech and transcription. Five months later, there is already a new version of the image model: from 2.5 to 2.6. Microsoft really seems to mean it with these in-house models. This doesn't feel like a side experiment next to OpenAI. It feels like a full model line that keeps improving fast.
Another thing that stands out: the same pattern shows up across the whole family, each time with a Flash version. MAI-Code-1-Flash, MAI-Voice-2-Flash and now MAI-Image-2.6-Flash. Microsoft is clearly choosing speed and price here, and it talks about "price per Elo" instead of just the best possible quality. Not necessarily the smartest model, but good enough at the lowest cost per call. For pipelines with lots of volume, that is what counts in the end.
For us as a Microsoft partner, this is interesting. Anyone building in Foundry can choose between a pricier frontier model and a cheaper Microsoft model in the same environment, with the same Entra sign-in and inside their own tenant. Picking a model becomes more and more a cost choice for each task, instead of a choice of provider.
My expectation: Microsoft will push the MAI models more and more in Foundry and Copilot. Less reliance on OpenAI, and more control over its own margins.
Google's new Gemini business hub links Workspace, Shopify and Xero, adds brand kits and finance charts, but small firms should weigh privacy and lock-in.
NVIDIA's Cosmos3-H-Dreams surgical simulator is visually impressive but lacks the physical validation and clinical testing surgeons require.
Mistral previews Mistral Large 4, a 1T-parameter multimodal MoE model featuring open weights, strong agentic performance, and EU data sovereignty.