Microsoft AI, Microsoft's artificial intelligence division, announced MAI Code 2 and a faster variant, MAI Code 2 Flash (Flash below), on November 17, 2026. Both models are built for everyday enterprise coding work and, as of this writing, are available in GitHub Copilot. The company is pitching them on cost, saying they deliver performance close to leading rival models at a fraction of the price.
MAI Code 2 Handles Production Code, While Flash Takes On High-Volume Routine Work
Microsoft lists MAI Code 2's strengths as taking code to production, deep debugging, multi-file implementation, repo-wide reasoning, and UI work. The company says it also tuned the model for dashboards, visualizations, and visual coding.
Flash, by contrast, is designed for high-volume tasks such as code explanations, targeted edits, routine fixes, and rapid iteration. Microsoft also states plainly that both models are "explicitly not yet tuned" for harder problems such as fundamental science research and very long-horizon agentic work.
Microsoft Says MAI Code 2 Reaches 98% of GPT-6.1 Sol Across Four Benchmarks
Microsoft's comparison is shown in the table below. The MAI Code 2 figures are an average over four benchmarks targeting the everyday enterprise coding tasks the company sees in GitHub Copilot: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1. Flash was compared across three benchmarks focused on day-to-day, high-volume tasks.
| Model | Compared with | Performance | Cost |
|---|---|---|---|
| MAI Code 2 | GPT-6.1 Sol | 98% | 40% |
| MAI Code 2 | Claude Opus 5.5 | 96% | 5% |
| MAI Code 2 Flash | GPT-6 Luna | Comparable | 65% |
| MAI Code 2 Flash | Claude Haiku 5.5 | Comparable | 20% |
Performance and cost in the table are expressed relative to the comparison model at 100%. Separately, in blind evaluations of generated frontend apps, 258 reviewers preferred MAI Code 2 over GPT-6 Sol 64% of the time, according to Microsoft.
The test conditions matter, however. All models were run at maximum reasoning effort, and three of the benchmarks, including DeepSWE 1.1, were run on MSBench, Microsoft's internal evaluation platform, using the GitHub Copilot CLI. The costs for MAI Code 2 and Flash were also calculated at a 50% promotional price. Microsoft itself notes that results vary by workload, configuration, reasoning effort, and pricing assumptions.
A Derivative Model, MAI Work Flash, Runs Behind Microsoft's Office Apps
Microsoft also described MAI Work Flash, a model optimized from Flash. It is not offered as a standalone model; instead, it serves multiple Microsoft products across Microsoft 365 and Fabric IQ.
In internal evaluations, Microsoft says MAI Work Flash performed similarly to GPT-6.1 Sol on five Word benchmarks while using about 42% fewer input tokens per call in production. In Teams, it beat GPT-6 Luna at tasks such as tracking decisions, using up to 28% fewer tokens on instruction following and tool calling. In Excel and PowerPoint, the company says it is on par with GPT-6 Luna.
Available in GitHub Copilot, With Foundry and Third-Party Services Next
As of this writing, the models are available as follows.
- MAI Code 2: GitHub Copilot
- MAI Code 2 Flash: GitHub Copilot; Microsoft Foundry (limited-access preview)
Microsoft says the models are coming soon to OpenRouter, Vercel, Hermes, OpenCode, and Kilo. The announcement does not say which GitHub Copilot plans include the models, what the per-token prices are, or how long the promotional pricing will last.
Our Take
What stands out is that Microsoft is positioning its own models not as the most powerful option, but as good-enough performance at a low price. Much of what developers do in GitHub Copilot every day is routine work like explanations and small fixes, and if those costs can be cut to a fraction, the appeal to businesses should be significant.
That said, every comparison figure comes from Microsoft's own testing, and the costs assume a 50% discount. How large the gap will be at regular prices after the promotion ends is unknown as of this writing. Still, Microsoft deserves credit for spelling out where the models fall short, and the realistic approach is to start with everyday tasks and check the results on one's own code.
Summary
Microsoft AI announced MAI Code 2 and MAI Code 2 Flash, coding models available in GitHub Copilot. The news matters to developers who use GitHub Copilot and to companies looking to rein in AI costs. The models are set to expand to outside services such as OpenRouter, and the next thing to watch is whether official per-token pricing and plan-by-plan availability are disclosed.
This article is based on Microsoft AI's November 17, 2026 announcement, "MAI Code 2 and Code 2 Flash: Everyday coding at scalable prices."