
Microsoft has released MAI Code 2, an in-house coding model. It says the model delivers 98% of the performance of OpenAI’s GPT-6.1 Sol at 40% of the price.
The company presented it at its Ignite conference in San Francisco on Tuesday. New models for security and image editing launched alongside it.
Against Anthropic’s Claude Opus 5.5, Microsoft says MAI Code 2 gives 96% of the performance at 5% of the cost. The figures are averages across four coding benchmarks that Microsoft chose.
Both comparisons use a launch price. A footnote to Microsoft AI’s announcement says the cost is shown at a 50% promotional rate. That rate runs until 17 May 2027, GitHub’s pricing page shows. After that, input tokens rise from $0.50 to $1 per million. Output tokens rise from $2.50 to $5 per million.
MAI Code 2 and a smaller model, MAI Code 2 Flash, are available in GitHub Copilot. Microsoft says Flash matches OpenAI’s GPT-6 Luna at 65% of the price. It matches Anthropic’s Claude Haiku 5.5 at 20%.
Mustafa Suleyman, chief executive of Microsoft AI, led the launch on stage.
“Right now, token spend is going through the roof,” Suleyman said, describing what he hears from CIOs and CFOs.
Microsoft says both coding models are not yet tuned for harder problems. Those include fundamental science research and very long agentic work.
A frontier model of its own
Microsoft is also building a general-purpose frontier model, Suleyman told Axios. He said the company has reached state-of-the-art pre-training results. It is also making strong progress in reinforcement learning. He gave no release date.
“My team is building true superintelligence,” Suleyman told Axios.
He gave no target for how much Copilot traffic will move to Microsoft’s own models. He predicted significant adoption.
Security and image models
MAI Cyber 2 Flash, a security model, beats OpenAI’s GPT-6 Astra, Suleyman said. It does so at less than 2% of the cost per task. A larger MAI Cyber 2 is in preview inside MDASH, Microsoft’s agentic code scanner.
MAI Image 2.7 is now the top image editing model, Suleyman said. It costs a fifth of what its rivals charge.
Licensed weights
Microsoft is also launching MAI Licensed Weights. Trusted customers can take a Microsoft AI model and customise it with their own data. They then run it in their own Azure environment, with help from Microsoft’s Frontier Company team.
Suleyman told Axios the approach would “feel like open source” with more security and control.
“No one wants to pour their IQ into somebody else’s model,” Suleyman said on stage.
AT&T built a custom version of Code 2 Flash. It beats GPT-6.1 Sol at standard 5G network analysis and costs 50 times less, according to Suleyman.
Microsoft still sells OpenAI and Anthropic models. Satya Nadella announced OpenAI’s GPT-6.2 for Microsoft Foundry on the same stage. Earlier, Microsoft’s commercial chief pitched Copilot as cheaper than ChatGPT and Claude.
Microsoft cut back Claude Code licences for its own engineers in May over token costs. In September, Suleyman took on Anthropic over how it trains Claude.