Microsoft introduced two coding models at its Ignite conference on November 17, bringing MAI Code 2 and the smaller MAI Code 2 Flash to GitHub Copilot. The company says the larger model comes close to competing systems on a set of software-engineering benchmarks at a fraction of their cost. There is a significant condition: Microsoft’s calculation uses a 50% promotional price for its own models.
For developers, the immediate choice is between a larger model aimed at multi-file work and a faster one intended for routine edits. Both are available through Copilot, while only Flash is currently offered through Microsoft Foundry as a limited-access API preview, according to Microsoft’s announcement.
What the price comparison actually measures
Microsoft reports that MAI Code 2 achieved 98% of GPT-6.1 Sol’s performance at 40% of the cost, averaged across DeepSWE 1.1, Terminal-Bench 2.1, SWE-Atlas and FrontierCode v1.1. It also claims 96% of Claude Opus 5.5’s performance at 5% of the cost on the same group of tests.
Those are benchmark-specific comparisons, not a promise that a team’s bill will fall by those amounts. Microsoft ran the models at maximum reasoning effort; three benchmarks used its internal evaluation platform and GitHub Copilot CLI without subagents, while FrontierCode figures came from Artificial Analysis. The fine print states that MAI Code 2’s cost was calculated at a 50% promotional price. The Flash comparison also uses a 50% promotional price.
Microsoft does not give a general post-promotion price in that announcement. Teams should check the applicable current rate and Copilot plan or Foundry terms before forecasting spend. Actual cost also depends on task length, model configuration, retries and how much input and output an agent uses.
Code 2 versus Flash: where each fits
Microsoft positions Code 2 for deeper debugging, repository-wide reasoning and changes spanning multiple files. Flash targets high-volume work such as explaining code, making targeted edits and handling routine fixes. Both accept text and images, allowing a developer to provide a screenshot or mockup alongside a coding request.
The model card lists a roughly 1.05-million-token context window, including limits of 922,000 input tokens and 128,000 output tokens. A large context window is a capacity limit, not evidence that every repository can be understood accurately in one pass. The models also offer adjustable reasoning effort, which trades speed and cost against quality.
The training and evaluation setup is relevant to Copilot users: Microsoft says it trained the models using the Copilot CLI and VS Code agent environments. Its announcement also notes that they are not yet tuned for fundamental science research or very long-horizon agent work. Treat that as a reason to test long, autonomous workflows separately from shorter coding tasks.
How to try them and evaluate a switch
In a supported Copilot client, look for MAI Code 2 or MAI Code 2 Flash in the model picker. The model card lists VS Code and Copilot CLI as examples. Availability can depend on the user’s plan and organization settings; verify the options and charges in your own account.
For an application built on Foundry, Microsoft lists Flash as a limited-access preview. Its model card describes creating a deployment and calling the model through the Responses API with Azure authentication and service controls. The larger Code 2 model is not listed there as a Foundry API offering. Microsoft has also announced licensed weights for selected enterprise arrangements running in Azure; that is separate from a generally downloadable open-weight release.
A useful trial is to run the same representative, non-sensitive issues through the current model and each MAI option. Record whether tests pass, how much human correction is needed, elapsed time and total cost per accepted change. Include a multi-file bug, a small routine edit and a task involving the team’s actual language and framework. Microsoft notes that performance outside English and on less common languages and frameworks has not been extensively evaluated.
Keep review and permissions in place
Generated changes still need code review, tests and security scanning. Microsoft’s model card recommends sandboxing agent work, using least-privilege access and treating repository content and tool output as possible prompt-injection sources. Limit access to credentials and require approval for commands or infrastructure changes with real-world side effects.
Microsoft’s own evaluation found that the models needed human intervention for complex errors or environmental changes. An independent assessment cited in the card also found susceptibility to some image-based and multi-turn jailbreaks. Those findings do not establish how often a particular team will encounter a failure, but they are a practical reason to keep boundaries and review in place during a pilot.
Axios reports that Microsoft is pursuing specialized models as it continues work on a broader general-purpose system. For teams choosing a coding assistant today, the more immediate question is whether these models solve their own tasks reliably at the price they actually pay.