Microsoft's MAI-Thinking-1 goes toe to toe with Claude Sonnet 4.6

Microsoft dropped its first reasoning model at Build 2026 Day 2, and the benchmarks are worth paying attention to. MAI-Thinking-1, a 35-billion active parameter Mixture of Experts model (roughly 1 trillion total parameters), claims human preference parity with Anthropic’s Claude Sonnet 4.6 in blind evaluations. It also matches Claude Opus 4.6 on SWE-Bench Pro.

That is a lot of firepower for a model that costs less to run than anything in Opus or Sonnet’s weight class.

The key claims from the Microsoft AI announcement:

• 97.0% on AIME 2025, 94.5% on AIME 2026 for mathematical reasoning
• Matches Opus 4.6 on SWE-Bench Pro for agentic coding tasks
• Preferred over Sonnet 4.6 in blind side-by-side human evaluations across 1,276 tasks
• 256K context window with function calling and enterprise API compatibility
• Trained without distillation from third-party models and built on what Microsoft calls “clean and commercially licensed data”

The “no distillation” claim is the most interesting part. Microsoft is explicitly saying this model learned from scratch rather than copying behavior from existing frontier models. That is a shot at the entire industry practice of training smaller models on outputs from larger ones. Whether that claim holds up under independent scrutiny is another matter. Simon Willison noted that the technical paper’s training data section (page 80+) reveals the model was trained on “a crawl of the public web,” which puts it in the same licensing gray area as every other major LLM.

The architecture tells a clear story about Microsoft’s strategy. A sparse MoE with 35B active out of ~1T total parameters means Microsoft is optimizing for inference cost. You get frontier-class reasoning without paying frontier-class compute bills. This runs on Microsoft’s own Maia 200 silicon inside Azure, which means higher margins per query than routing through NVIDIA GPU instances.

The competitive picture. Sonnet 4.6 at $3/$15 per million tokens is currently the most popular enterprise API model. If MAI-Thinking-1 genuinely performs at Sonnet 4.6 level and Microsoft prices it below that through Azure AI Foundry, every Azure-heavy enterprise has a reason to switch. CNBC reported that the model is available in private preview, so widespread pricing data is not yet public.

MAI-Thinking-1 was announced alongside six other MAI models. MAI-Code-1-Flash (5B active parameters, rolling out to all GitHub Copilot plans) is the coding companion. MAI-Image-2.5, MAI-Voice-2, and MAI-Transcribe-1.5 round out the multimodal stack. The whole family runs on Microsoft-designed Maia 200 chips connected through Microsoft’s MRC networking protocol. Gizmodo framed the launch as Microsoft exploiting legal anxieties around AI data provenance to sell businesses on its “clean data” story.

The “hill-climbing machine” framing from Mustafa Suleyman’s team is worth noting. This is not positioned as a one-off model launch. The entire pipeline (data curation, reward modeling, RL environments, silicon co-design) is built to iterate. Each new model in the MAI family is supposed to climb higher on the same infrastructure. If the pipeline works as described, the gap between Microsoft’s in-house models and Anthropic/OpenAI should narrow with each release cycle.

ZDNET highlighted that Build 2026 is being called Microsoft’s “AI independence day” by observers. Between the Polaris coding model (Day 1), MAI-Thinking-1 (Day 2), the Maia 200 silicon running in production, and the cancellation of internal Claude Code licenses, the trajectory is unambiguous. Microsoft is building its own models, running them on its own chips, and pricing them to undercut the partners it spent $18 billion investing in.

The model is in private preview now. General availability timeline was not disclosed.