Nearly as good for a fraction of the price, Microsoft says of its in-house coding models MAI Code 2 and Code 2 Flash
At its Ignite conference Microsoft presented seven of its own AI models, and the two coding models are already in GitHub Copilot.
Microsoft has launched a series of in-house AI models at its Ignite conference, intended to make the company less dependent on OpenAI and Anthropic. TestingCatalog counts seven, from image generation to speech and security. In the spotlight are MAI Code 2, meant for heavier work such as debugging and changes spanning many files, and the lighter MAI Code 2 Flash for quick edits and routine fixes. Both are available right away in GitHub Copilot.
According to Microsoft's figures, reported by InfoWorld, MAI Code 2 averages 63.2 percent on four coding benchmarks, against 64.1 percent for GPT-6.1 Sol and 65.3 percent for Claude Opus 5.5. Its cost per task, at 0.31 dollars, is considerably lower than those rivals. Microsoft itself says the models are not tuned for scientific research or for very long autonomous agent tasks.
That comparison deserves a caveat, analysts warn in InfoWorld. The figures use a half-price promotional rate that runs until 17 May 2027; after that MAI Code 2 costs 1 dollar per million input tokens and 5 dollars per million output tokens. One analyst calculates that a 60 percent saving on model costs yields only about 6 percent if models make up a tenth of development costs, and advises companies to test the models on their own code.
The other new models include MAI Cyber 2 Flash, a security model coming exclusively to Microsoft's own security products from mid-December, according to TestingCatalog, and the MAI Image 2.7 image models, in preview on the Foundry platform. Microsoft also says its own Maia 200 AI chip is running its largest models in production.