Microsoft ships seven MAI models and Project Polaris, pulls Copilot off OpenAI's stack at Build 2026
Satya Nadella opened Microsoft Build 2026 in San Francisco on Tuesday with a model family Microsoft trained from scratch on its own Maia accelerators. MAI-Thinking-1, a 35-billion-parameter reasoning model, matches Claude Opus 4.6 on the SWE Bench Pro coding benchmark; MAI-Image-2.5 entered Arena.ai's image leaderboard at number three; and Project Polaris will replace GPT-4 Turbo as the default engine across every paid GitHub Copilot tier starting in August. The Copilot transition ends Microsoft's three-year dependence on OpenAI as the substrate for its most-shipped developer product.
Thurrott / Tech Times / Microsoft Foundry · 2026-06-02 · Source ↗ ·
Microsoft ships seven MAI models and Project Polaris, pulls Copilot off OpenAI's stack at Build 2026
Satya Nadella opened Microsoft Build 2026 in San Francisco on Tuesday with a model family Microsoft trained from scratch on its own Maia accelerators. MAI-Thinking-1, a 35-billion-parameter reasoning model, matches Claude Opus 4.6 on the SWE Bench Pro coding benchmark; MAI-Image-2.5 entered Arena.ai's image leaderboard at number three; and Project Polaris will replace GPT-4 Turbo as the default engine across every paid GitHub Copilot tier starting in August. The Copilot transition ends Microsoft's three-year dependence on OpenAI as the substrate for its most-shipped developer product.
Thurrott / Tech Times / Microsoft Foundry · 2026-06-02 · Source ↗ ·
Satya Nadella opened the Microsoft Build 2026 keynote at Fort Mason Center in San Francisco at 9:30 a.m. Pacific on Tuesday and spent the first thirty minutes of the program walking the audience through a frontier-model family Microsoft trained itself. The MAI line, formally introduced through the Microsoft Foundry blog the same morning, contains seven models that together cover reasoning, code, image generation, image editing, speech transcription, speech synthesis, and the Copilot default engine. Microsoft describes each model as "trained from the ground up on clean data, without distillation from third-party models," a phrase the launch material returned to repeatedly. The strategic message landed before the technical detail: Microsoft is no longer content to be the best enterprise distributor of someone else's AI.
The headline release is MAI-Thinking-1, a 35-billion-parameter reasoning model that Microsoft says matches Anthropic's Claude Opus 4.6, released in February, on the SWE Bench Pro coding benchmark. The model is in private preview through Microsoft Foundry to a select group of early partners. Microsoft characterizes Thinking-1 as a "medium-sized" model designed for high efficiency at a low per-token cost, which is the wedge it wants to push against frontier-priced offerings from OpenAI, Anthropic and Google. The company has not yet disclosed token pricing or context length, which will follow the public-preview window.
Polaris and the end of the GPT-4 era inside Copilot
Project Polaris is the second strategic release and the one with the most immediate revenue consequence. Polaris is a mixture-of-experts coding model with language-specific sub-modules that Microsoft says outperforms GPT-4 Turbo on HumanEval and MBPP, the two benchmark suites that have defined coding-model comparison since 2023. Microsoft has not published independent verification for those numbers, and the Tech Times writeup of the keynote explicitly notes the benchmark figures are Microsoft's own. Polaris runs on Microsoft's custom Maia AI accelerators inside Azure. Beginning in August 2026, Polaris will replace GPT-4 Turbo as the default engine for every paid GitHub Copilot tier. An optional three-month fallback to GPT-4 will be available before the August deadline.
GitHub CEO Thomas Dohmke described the Polaris transition as "the biggest change to Copilot since launch," a phrase that landed against an audience that remembers Copilot's October 2021 debut. The default-model swap is the single largest concrete commitment of the morning, and it ends the substrate dependency that has defined Microsoft's developer-tooling product since the original Codex partnership. Copilot users will be able to opt back to OpenAI or Anthropic models via the model-picker, but the default is the substrate that drives ninety-plus percent of usage in practice.
The image, voice, transcription, and code family
MAI-Image-2.5 entered Arena.ai's leaderboard at number three for image-generation model families, beating Google's Nano Banana Pro on ELO. The model ships in two variants, a high-quality version and a faster MAI-Image-2.5e variant, and accepts image uploads, which opens the model to image-editing as well as generation. Pricing in the Foundry Model Catalog starts at $5 per million tokens for text input, $8 per million tokens for image input, and $47 per million tokens for image output. MAI-Image-2.5 introduces identity and character consistency, preserving recognizable faces, hair, clothing, and full-body identity across stylization, pose, and layout changes, which is the feature that drove the Arena rating.
MAI-Voice-2 is the second-generation speech synthesis model, paired with a faster Flash variant. MAI-Transcribe-1.5 is the new speech-recognition model that Microsoft says is "five times faster than competing models" and offers "state-of-the-art accuracy" across 43 languages. MAI-Code-1 is the production coding model that ships directly into GitHub Copilot and Visual Studio Code, separate from Polaris (which is the orchestration default). The launch lineup also touches the third-party catalogue: Microsoft announced that Fireworks AI is now generally available on Microsoft Foundry with enterprise SLAs and SOC2 readiness, giving developers a single Azure endpoint into more than twenty open frontier models including DeepSeek, Kimi, GLM, gpt-oss, and MiniMax.
"MAI-Thinking-1 was trained from the ground up on clean data, without distillation from third-party models."— Microsoft Foundry launch material, June 2, 2026
The strategic frame
The MAI launch is the moment Microsoft completes a three-year repositioning. The company entered 2023 as the distribution channel for OpenAI's frontier models; it exits Build 2026 as a model maker that owns the runtime (Foundry), the silicon (Maia accelerators in Azure datacenters and the RTX Spark superchip on the client), the agentic SDK (Microsoft Agent Framework), and the productivity surface (Copilot across Office and GitHub). The OpenAI partnership remains, and OpenAI's Codex models continue to be available through Copilot and Foundry, but the default has shifted. Microsoft no longer depends on OpenAI's frontier-model release cadence to keep Copilot competitive.
The Polaris transition specifically tracks a pattern that began in March when GitHub Copilot CLI reached general availability with multi-model support and accelerated through Cursor's, Zed's, and Anthropic's Claude Code product entries during Q1. The agentic-coding category has compressed from a single OpenAI-dominated market into a multi-model contest in which Microsoft, Anthropic, and OpenAI each ship competitive offerings. By owning the default, Microsoft captures the margin tier that matters most in a metered-pricing environment, the tier that ships before the user changes any settings.
"The biggest change to Copilot since launch."— Thomas Dohmke, GitHub CEO, on the Polaris default transition
What the developer audience saw next
The MAI announcement opened a keynote that ran another two hours and covered Project Solara, the new agent-first device platform; Azure Agent Mesh, the federated agent-orchestration service; Aion 1.0, the new on-device Windows model line; WSL 3, with paravirtualized GPU and NPU passthrough; the standalone GitHub Copilot desktop app; Microsoft Execution Containers for agent isolation; Coreutils for Windows; the Arm Development Kit 2026; and the Surface RTX Spark Dev Box. The MAI family is the substrate. The rest of the keynote is what Microsoft built on top. By 12:30 p.m. Eastern when Nadella handed the stage to GitHub's session, Microsoft had announced more shippable AI surface area in a single morning than at any developer conference since the 2017 Azure Stack debut.