Microsoft: MAI models beat Anthropic and Google rivals

Microsoft introduced seven new MAI models at its Build developer conference on Tuesday, and presented test results it says show its models outperforming rival systems. The company reported that independent evaluators preferred MAI-Thinking-1 over Anthropic’s Claude Sonnet 4.6 in blind comparisons and that MAI-Image-2.5 ranked higher than Google’s Nano Banana 2 on image-editing leaderboards.

The company described MAI-Thinking-1 as a flagship reasoning model. Microsoft reported the model scored 97% on AIME 2025, a benchmark for advanced problem solving and reasoning. Results on the SWE Bench Pro coding test were presented as comparable to Anthropic’s Opus 4.6. Microsoft said independent evaluators conducted the blind comparisons but did not publish full raw data for every test at the event.

The MAI family includes models for reasoning, coding, image editing, transcription and voice. Microsoft introduced MAI-Code-1-Flash for software development use in GitHub Copilot and Visual Studio Code. MAI-Image-2.5 and a Flash variant were shown for image-editing tasks. MAI-Transcribe-1.5 supports transcription in 43 languages. MAI-Voice-2 is a speech-generation model available in 15 languages and can adapt to a speaker from a short audio sample. Microsoft also reported MAI “delivered the highest win rate, outperforming GPT-5.5 on quality, while being 10x lower on cost.”

Microsoft described the models as built from scratch with a clean data lineage and optimized for efficiency. The company said the lineup is intended for integration into developer tools and business applications and emphasized cost and performance as factors for adoption.

Mustafa Suleyman, chief executive of Microsoft AI, posted on X that he was “Super excited to announce seven new world-class MAI models today” and characterized the release as a family of models designed to work together. In a company blog post he wrote that the compute used to train frontier models has grown by a factor of one trillion and projected another thousand-fold increase in the next three years.

The launch comes as other AI developers have released new models and updates in recent weeks. Microsoft positioned the MAI lineup as part of its effort to build proprietary frontier models while continuing its existing partnership with OpenAI.

Microsoft provided benchmark summaries and selective comparison data at Build. The company did not release complete raw data for all evaluations at the event.