Microsoft MAI-Thinking-1 Hits 97% on AIME — First Reasoning Model Built Without Distillation, Matches Opus 4.6 on SWE-Bench Pro
Microsoft shipped MAI-Thinking-1 at Build 2026 on June 2 — its first large language model built entirely in-house and the first model from its new AI model development pipeline it calls a “hill-climbing machine.”
The model is a sparse Mixture-of-Experts architecture with 35 billion active parameters and roughly 1 trillion total parameters. It is currently in private preview on Microsoft Foundry.
What It Claims
| Benchmark | MAI-Thinking-1 |
|---|---|
| AIME 2025 | 97.0% |
| AIME 2026 | 94.5% |
| LiveCodeBench v6 | 87.7% |
| SWE-Bench Pro | 52.8% |
Microsoft says it matches Claude Opus 4.6 on SWE-Bench Pro. The positioning is precise: Opus 4.6 scores 52.5% on SWE-Bench Pro. The claim is parity at substantially lower cost.
In blind side-by-side evaluations, Microsoft says users preferred MAI-Thinking-1 over Sonnet 4.6. That is a weaker claim than matching Opus 4.6 — Sonnet is the mid-tier model — but it suggests the model holds up in general instruction following, not just structured benchmarks.
No Distillation
The training method is the notable detail. Microsoft explicitly says it trained from the ground up on “enterprise grade, clean and commercially licensed data, without distillation from third-party models.”
Distillation — training a smaller model on the outputs of a larger one — has been the standard shortcut for reaching frontier performance on a budget. DeepSeek-R1 used it. OpenAI’s o-series models benefited from it. Musk testified under oath that xAI distilled OpenAI models to build Grok. Microsoft is drawing a sharp line here, partly for legal defensibility with enterprise customers and partly to signal that it has built a real training pipeline, not an imitation.
The base model trained on 30 trillion tokens, mostly human-generated. Post-training used reinforcement learning across math, coding, tool use, helpfulness, and safety.
The Hill-Climbing Machine
Mustafa Suleyman described the broader ambition: Microsoft now has what he calls a “hill-climbing machine” — a connected process that improves data, training setup, rewards, safety tests, and evaluations as a single loop. MAI-Thinking-1 is the first model out of that loop.
The architecture fits the strategy. MoE means 35B active parameters handle inference while 1T total parameters provide the knowledge capacity. The result is a smaller inference footprint than a dense 70B model, with compute costs that make high-volume enterprise deployment viable. Microsoft’s pitch is not “our best model ever” — it is “our frontier-class model that enterprises can actually run every day.”
Context in the Market
At 52.8% SWE-Bench Pro, MAI-Thinking-1 sits significantly below the current frontier. For reference:
- Claude Opus 4.8: 69.2% SWE-Bench Pro
- Claude Opus 4.7: 64.3% SWE-Bench Pro
- Qwen3.7-Max: ~60.6% SWE-Bench Pro
- GPT-5.5: 58.6% SWE-Bench Pro
- MAI-Thinking-1: 52.8% SWE-Bench Pro (Opus 4.6 level)
That is 2026’s mid-tier, not the frontier. But for enterprises running coding workflows at scale — where cost per task matters more than marginal capability — a model that performs like Opus 4.6 at lower cost and with clean IP provenance is a real offer.
Pricing for MAI-Thinking-1 has not been announced. MAI-Image-2.5, released alongside it, is priced at $5/M input (text) and $47/M output (image). MAI-Code-1-Flash, the coding-optimized model also at Build 2026, was previously detailed.
MAI-Thinking-1 joins a Build 2026 lineup that also included MAI-Code-1-Flash, MAI-Image-2.5, MAI-Image-2.5-Flash, MAI-Transcribe-1.5, and Scout, a personal AI agent. Microsoft is now fielding every model category in-house.