Photo: Microsoft
SAN FRANCISCO, November 17, 2026 — Microsoft has launched a batch of in-house AI models at its Ignite conference that it says match or beat rivals from OpenAI and Anthropic at a fraction of the cost, including a cybersecurity model it claims outperforms GPT-6 Astra on a vulnerability-finding test for less than 2 per cent of the cost per task.
The models, from Microsoft AI (MAI), span coding, security, image editing, voice and transcription. In its announcement, Microsoft listed seven new models; Axios described six specialised models, so counts differ slightly between the two.
The security model is the eye-catcher. Microsoft said MAI Cyber 2 Flash scored 92.3 per cent on the CyberGym-E2E benchmark for real-world vulnerability discovery, against 91.5 per cent for GPT-6 Astra, at less than 2 per cent of the cost per task. Its list price is US$0.10 per million input tokens and US$0.50 per million output tokens, with a one-million-token context window, four times larger than its predecessor. Microsoft said it excluded 92 of the benchmark’s 905 tasks because they overlapped with its training data. The model will come only through Microsoft’s own security products, starting from mid-December for eligible customers.
For developers, MAI Code 2 is available in GitHub Copilot. Microsoft says it reaches 98 per cent of GPT-6.1 Sol’s coding performance at about 40 per cent of the price, and 96 per cent of Anthropic’s Claude Opus 5.5 at 5 per cent of the price. A smaller MAI Code 2 Flash is aimed at high-volume routine coding. Microsoft also launched MAI Image 2.7 for image editing, MAI Realtime 1 for voice and a transcription model for healthcare.
The claims come with caveats. Axios noted that Microsoft’s comparisons rely on selected benchmarks and that some pricing advantages reflect promotional discounts. Specialist models also trade breadth for cost: one tuned for everyday coding will not match general models at advanced maths or science.
Microsoft AI chief executive Mustafa Suleyman told Axios the company was also making strong progress on its own frontier general-purpose model, but declined to give a release date. “My team is building true superintelligence,” he said.
Why it matters: Microsoft has long relied on OpenAI for its most advanced AI. Cheap in-house models running inside GitHub Copilot and Microsoft’s security tools could move a big share of that work onto Microsoft’s own technology, and push prices down across the market.
What to watch next: independent tests of the MAI models, how much Copilot traffic shifts to them, and whether Microsoft sets a date for its own frontier model.
Sources: Microsoft AI; Axios.

