THE CRUNCH
Microsoft's AI division has unveiled seven new in-house models at its Ignite conference, released in a single batch just as it did at Build in June. The lineup covers code, security, images, speech and routing, and Microsoft is positioning it primarily on price-to-performance, with OpenAI's models as the main yardstick. The company points to rapidly rising token spending among customers that want to run AI agents around the clock.
The headline claim is MAI Code 2, which Microsoft says averages 98 percent of GPT-6.1 Sol's performance across four benchmarks at 40 percent of the cost, and reaches 96 percent of Claude Opus 5.5's performance at 5 percent of the cost. In blind comparisons of frontend work such as UIs and data visualisations, human evaluators preferred its output 64 percent of the time over GPT-6 Sol, Microsoft says. In GitHub Copilot it costs $0.50 per million input tokens and $2.50 per million output tokens at a promotional price through May 17, 2027, then $1 and $5.
On security, MAI Cyber 2 Flash solves 92.3 percent of tasks in the CyberGym-E2E vulnerability benchmark, Microsoft says, edging out GPT-6 Astra at 91.5 percent, a model that needs twice the reasoning budget. Per task it works out at $0.45 instead of $34.38. Microsoft also discloses that during evaluation it found 10 percent of the roughly 900 CyberGym tasks drew on the same open-source projects and vulnerability reports as parts of its own training data; it says reference solutions never entered training, and removed those tasks for all models anyway.
Microsoft is candid that the code models target everyday enterprise coding tasks and are not yet optimised for harder work such as foundational research or long-horizon agentic workflows. A comparison with GPT-6 Cyber is also missing so far, as that model is available only to selected customers, and agentic benchmarks with tools still need extra safety measures. Microsoft plans to publish those results in the coming weeks.


