Microsoft has introduced MAI Code 2 and MAI Code 2 Flash, two in-house coding models that can now be selected in GitHub Copilot. The models come from the Microsoft AI (MAI) team and, according to the model card, were pre-trained and post-trained entirely by Microsoft. The pitch: close to the quality of the current top models from OpenAI and Anthropic at a fraction of the cost per task. A closer look, however, reveals a discounted price comparison, several limitations and a model card with contradictory figures.
MAI Code 2: What Microsoft announced
The Microsoft AI announcement was published on November 17, 2026, at the Ignite conference. MAI Code 2 is the larger model and, according to Microsoft, is strongest at debugging, multi-file changes, reasoning across whole repositories and UI code. MAI Code 2 Flash is the smaller, cheaper variant for code explanations, targeted edits, routine fixes and rapid iteration.
The two models are part of a set of seven new MAI models Microsoft presented on the same day. They include MAI Cyber 2 Flash for security tasks, the image model MAI Image 2.7, the voice model MAI Realtime 1 and the decision model Decision 1.1. Microsoft also said its Maia 200 AI accelerator is now running its largest models in production.
One limitation is stated by Microsoft itself: the models are explicitly not yet tuned for harder problems such as fundamental science research or very long-horizon agentic work. The focus is on the coding work companies do every day.
The benchmarks according to Microsoft
Microsoft compares MAI Code 2 with OpenAI’s GPT-6.1 Sol and Anthropic’s Claude Opus 5.5. The basis is the average of four benchmarks: DeepSWE 1.1, Terminal-Bench 2.1, SWE-Atlas and FrontierCode v1.1. By that measure, MAI Code 2 delivers about 98 percent of GPT-6.1 Sol’s performance at about 40 percent of the cost, and 96 percent of Claude Opus 5.5’s performance at 5 percent of the cost.
InfoWorld compiled the individual results from Microsoft’s figures:
- Average: MAI Code 2 scores 63.2 percent, GPT-6.1 Sol 64.1 percent and Claude Opus 5.5 65.3 percent.
- DeepSWE 1.1: 73.2 percent versus 74.6 (GPT-6.1 Sol) and 69 (Opus 5.5).
- Terminal-Bench 2.1: 84.6 percent versus 85 and 78.7.
- SWE-Atlas: 43.1 percent versus 47.4 and 57.5.
- FrontierCode v1.1: 52 percent versus 49.4 and 56.
- Cost per task: 0.309 dollars for MAI Code 2, 0.761 dollars for GPT-6.1 Sol and 6.306 dollars for Claude Opus 5.5.
For MAI Code 2 Flash, Microsoft claims performance comparable to GPT-6 Luna and Claude Haiku 5.5 on three benchmarks, at 65 and 20 percent of their cost respectively. According to InfoWorld, the cost per task is 0.052 dollars versus 0.080 dollars (GPT-6 Luna) and 0.319 dollars (Haiku 5.5).
There is also a blind study on frontend work: across user interfaces, games, data visualizations, 3D scenes and SVG graphics, developers preferred MAI Code 2’s results over GPT-6 Sol 64 percent of the time. According to the footnote, 258 annotators rated both looks and functionality.
All figures come from Microsoft. The tests were run at maximum reasoning effort on MSBench, Microsoft’s internal evaluation platform, using the Copilot CLI as the harness and without subagents. There are no independent replications yet.
The catch: comparing at a promotional price
The cost advantage is smaller than the headline suggests. A footnote in the announcement states that the costs for MAI Code 2 and Flash were calculated at a 50 percent promotional price. According to the Copilot pricing page cited by InfoWorld, MAI Code 2 costs 0.50 dollars per million input tokens, 0.025 dollars per million cached input tokens and 2.50 dollars per million output tokens until May 17, 2027. After that, prices double to 1.00, 0.05 and 5.00 dollars. Flash starts at 0.05, 0.005 and 0.25 dollars, later rising to 0.10, 0.01 and 0.50 dollars.
Once the promotion ends, MAI Code 2 would therefore land at roughly 80 percent of GPT-6.1 Sol’s cost rather than 40 percent, assuming other prices stay the same. For Flash, the advantage over GPT-6 Luna would disappear entirely on paper: 0.052 dollars per task would become about 0.104 dollars, compared with 0.080 dollars for GPT-6 Luna. Analyst Sanchit Vir Gogia of Greyhound Research also told InfoWorld that Microsoft compares the spend of a benchmark run, not a published price per token. How many tokens a model needs for a task depends heavily on the workload.
What the model card reveals
The model card provides technical details but is not consistent on every point. Both models are decoder-only transformers with Mixture-of-Experts layers and an integrated vision encoder. They accept text and images, the context window is about 1.05 million tokens and output is capped at 128,000 tokens. Their knowledge runs to August 2026. The weights are not released; the models are only available as a hosted service on Azure.
On size, the card contradicts itself: the summary table lists MAI Code 2 at “50B-100B (80B active)”, while the body text says 77 billion active and about 2.8 trillion total parameters. For Flash, the table lists 5 to 15 billion parameters, the text 12 billion active and about 398 billion total. Microsoft does not say which figure is correct.
The limitations Microsoft lists itself are also worth noting:
- At launch, the models are optimized for English-language code production only.
- In long agentic sessions, they can repeat tool calls, drift from instructions or stop before fully verifying their work.
- On alignment, Microsoft rates both models as similar to GPT-5.6 Sol but less aligned than GPT-6.
- On SWE-Atlas, the small Flash model scores slightly higher than the large model in the card, at 44.8 versus 43.1 percent.
For biological and cyber risks, Microsoft uses a two-tier classifier system. A fast filter detects relevant content, and a second reasoning model blocks responses above a risk threshold. The card cites chained exploit development as an example. Repeated violations are reviewed by humans and can lead to service suspension.
Availability in GitHub Copilot
According to the GitHub changelog, MAI Code 2 is available in the Copilot Pro, Pro+, Max, Business and Enterprise plans. MAI Code 2 Flash is also available in Copilot Free and Student. Billing is usage-based. The models can be selected in VS Code, Visual Studio, the Copilot CLI, the cloud agent, on github.com, in the mobile apps and in JetBrains IDEs, Xcode and Eclipse.
Important for companies: in Copilot Business and Enterprise, new models are enabled automatically by default unless administrators have turned off the global default or explicitly disabled the model. Teams that want to vet models internally before use should check the model policy in the Copilot settings.
MAI Code 2 Flash is also available as a limited-access preview in Microsoft Foundry, where it can be called through an OpenAI-compatible Responses API. Microsoft says more platforms, including OpenRouter and Vercel, will follow soon.
Assessment
MAI Code 2 is above all a strategic signal. Alongside its partnership with OpenAI, Microsoft is building its own models for its most important developer product and is competing on price rather than peak performance. For teams already using Copilot, it is one more option in the model picker that can be tested against their own tasks. But the advertised cost advantage only applies in full until May 2027, and the benchmarks come exclusively from Microsoft.
A comparison with the newly released GPT-6.2 Astra and Sol is worthwhile, since OpenAI is also selling those models on efficiency rather than raw capability. For a sense of how such models hold up in everyday editor work, our review of the Orca AI editor shows what matters in practice: traceable changes that you check yourself before merging.




