Intelligence Report

Microsoft's MAI launches seven in-house models priced to undercut OpenAI and Anthropic

At Ignite, Microsoft's AI division introduced models for coding, security, images, speech and routing, built to lower the cost of always-on enterprise agents.

INSIGHTEX Editorial Desk

2 min read

Microsoft's MAI launches seven in-house models priced to undercut OpenAI and AnthropicIntelligence Report

What happened

Microsoft's AI division, MAI, introduced seven specialized models at its Ignite conference, spanning code generation, cybersecurity, picture editing, voice and workflow routing, The Decoder reported. Microsoft is selling the batch mainly on price-to-performance for businesses that keep AI agents running continuously. It points to fast-growing token bills among those customers and benchmarks largely against partner OpenAI, as well as Anthropic.

Microsoft says MAI Code 2 reaches an average of 98 percent of GPT-6.1 Sol's results on four benchmarks while costing 40 percent as much. It says the model reaches 96 percent of Claude Opus 5.5's results at 5 percent of the cost. Those cost figures assume a 50 percent promotional discount. In GitHub Copilot, the model costs $0.50 for each million input tokens and $2.50 for each million output tokens until May 17, 2027, then rises to $1 and $5.

According to Microsoft, MAI Cyber 2 Flash solved 92.3 percent of CyberGym-E2E tasks, compared with 91.5 percent for GPT-6 Astra, at about $0.45 per task versus $34.38. Other releases include MAI Image 2.7, the speech-to-speech model MAI Realtime 1, and Decision 1.1, a routing model built on Google's Gemma 4. A planned MAI Licensed Weights offering will let enterprises tune selected models on their own data inside Azure. Its launch date and pricing are not yet set.

Context

Microsoft followed the approach it used at Build in June, releasing its models together rather than across the year. Bing Image Creator has relied on MAI Image as its standard model since June. On MDASH, Microsoft's security platform, close to 40 percent of scans are already run by MAI Cyber models. The earlier version of Decision was built on Qwen3.5-9B. Microsoft now runs its large models on Maia 200, the second generation of its AI accelerator.

Astra profile

GitHub profile

SourceThe Decoder

The briefing

The week in tech, in one email.

The stories worth your time, every week. Unsubscribe anytime.