Microsoft announced at the June Build conference that it is rolling out seven new AI models, covering reasoning, code, images, transcription, and speech. In four core programming benchmark tests, MAI-Code-1-Flash outperformed Claude Haiku 4.5 across the board, and on SWE-Bench Verified, token usage was reduced by up to 60%.
List of 7 models: full coverage across five major areas
At the Build conference, Microsoft has publicly released six models with complete names: MAI Thinking-1 (reasoning), MAI Code-1-Flash (code), MAI Image-2.5 (images), MAI Image-2.5-Flash (a streamlined image variant), MAI Transcribe-1.5 (transcription), and one speech model. The full name of the 7th model has not been fully disclosed in existing coverage. Microsoft said its flagship reasoning model MAI-Thinking-1 is competitive with peer models on STEM reasoning and coding tasks, and did not provide specific benchmark data.
MAI-Code-1-Flash vs Claude Haiku 4.5: data from four benchmark tests
MAI-Code-1-Flash had higher pass rates than Claude Haiku 4.5 in the following four benchmark tests, while using fewer tokens:
SWE-Bench Pro: 51.2% for MAI-Code-1-Flash vs 35.2% for Claude Haiku 4.5, leading by 16 points
SWE-Bench Verified: token usage for solving complex problems reduced by up to 60%
IF Bench (precise instruction following): MAI-Code-1-Flash leads by 28.9 points
Advanced IF: MAI-Code-1-Flash leads by 14.5 points
Microsoft’s official documentation notes that MAI-Code-1-Flash’s accuracy in core adversarial categories such as the Einstellung trap is still below 50%, representing identified areas for further improvement.
Current availability range of MAI-Code-1-Flash: no extra setup needed for individual users
MAI-Code-1-Flash is now available in the personal user version of VS Code’s GitHub Copilot. It can be used directly through the model selector, or Copilot can automatically route it via the selector. The model uses adaptive resolution-length control: it stays concise for simple requests and allocates more reasoning resources to complex tasks. The rollout schedule for the enterprise edition and other user groups for MAI-Code-1-Flash has not been announced yet.
FAQ
What are the seven new models released by Microsoft Build 2026?
The six models with fully disclosed names include MAI Thinking-1, MAI Code-1-Flash, MAI Image-2.5, MAI Image-2.5-Flash, MAI Transcribe-1.5, and one speech-related model; the full name of the 7th model has not been fully revealed in existing official reports.
How is the 60% token reduction for MAI-Code-1-Flash achieved?
Microsoft said the model is trained using adaptive resolution-length control, automatically adjusting the depth of responses according to task complexity. Microsoft’s SWE-Bench Verified test results show that when solving complex problems, token usage can be reduced by up to 60%.
Which users does MAI-Code-1-Flash currently support, and what is the enterprise rollout timeline?
As of June 2026, MAI-Code-1-Flash has been made available to VS Code GitHub Copilot individual users, with no additional setup required. Microsoft has not announced the rollout schedule for the enterprise edition or other user groups.
