Summary
- Microsoft announced that its MAI-Thinking-1 model excelled in blind evaluations compared to Anthropic's Claude Sonnet 4.6 and matched Claude Opus 4.6 in a coding benchmark.
- The firm's MAI-Image-2.5 models reportedly outperformed Google's Nano Banana 2 in image-editing tests.
- This launch represents Microsoft's most ambitious initiative to create proprietary frontier AI models, alongside its collaboration with OpenAI.
During the opening day of the annual Microsoft Build conference, the technology giant introduced seven new AI models, asserting that these surpassed Anthropic's Claude Sonnet 4.6 and Google's Nano Banana 2 in blind assessments and image-editing standards.
This announcement reflects Microsoft's goal to position itself as a leading AI developer rather than just a major supporter and infrastructure provider for OpenAI.
Microsoft AI CEO Mustafa Suleyman expressed his excitement on X, stating, "We are thrilled to unveil seven new world-class MAI models today. They signify a new era in AI that empowers users and keeps them at the forefront of technology."
At the core of this release is MAI-Thinking-1, which Microsoft identifies as its primary text foundation model.
Seven new models launching at Build: let’s go!
Reasoning. Code. Image. Transcribe. Voice.Developed from the ground up with a clean data lineage, designed for efficiency, and functioning cohesively as a family of models.
Thread 🧵 #MSBuild pic.twitter.com/g3WQIcIQ24
— Microsoft AI (@MicrosoftAI) June 2, 2026
Suleyman noted that MAI-Thinking-1 received higher ratings than Anthropic's Claude Sonnet 4.6 in independent blind tests. He mentioned that the model achieved a score of 97% on AIME 2025, a benchmark assessing advanced reasoning and problem-solving abilities.
He also stated that the SWE Bench Pro results position it "on par with Opus 4.6 on one of the most challenging coding benchmarks."
Alongside MAI-Thinking-1, Microsoft introduced MAI-Code-1-Flash, a streamlined coding model designed for GitHub Copilot and Visual Studio Code; MAI-Image-2.5 and its Flash variant, both of which reportedly exceed Google's Nano Banana Pro in image-editing capabilities; MAI Transcribe-1.5, supporting transcription in 43 languages; and MAI-Voice-2, a speech generation model that can produce natural-sounding voices in 15 languages, adjusting to a speaker's voice with a brief audio sample.
“We are in an extraordinary technological era. The computational power used to train frontier models has surged by a trillion-fold,” Suleyman elaborated in a separate blog post regarding the new models. “We anticipate another thousand-fold increase in the next three years, leading to enhanced capabilities and a continued rollout of increasingly effective AI.”
This announcement arrives amid escalating competition among top AI developers.
Recently, Anthropic launched its newest flagship model, Opus 4.8, which the company claims is faster and smarter in benchmark testing and includes a range of new features. Additionally, Anthropic revealed an expansion of its Project Glasswing, allowing 150 companies to access its new cybersecurity-focused Mythos model.
In May, Google introduced Gemini Omni at Google I/O, a multimodal AI model that integrates Gemini with its Veo, Nano Banana, and Genie media-generation models, in addition to Gemini Spark, a cloud-based AI agent designed to handle tasks across applications and workflows for users.
Microsoft's latest model launch indicates a strategic shift towards developing proprietary AI solutions, moving away from its historical dependence on OpenAI technology, asserting that MAI “achieved the highest win rate, outperforming GPT-5.5 on quality while costing 10 times less.”
“Developers and businesses have been demanding AI that operates on their terms and under their control,” Suleyman stated. “We view this as a significant step toward fulfilling that demand.”