MAI-Thinking-1 ranks first on 2 of the 20 benchmarks it has scores on.
Microsoft Jun 2026 Flagship
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| TruthfulQA | 88% | #1 of 28 | Model card | |
| AdvancedIF | 85% | #1 of 6 | Model card | |
| MMLU-Pro | 85% | #29 of 151 | Model card | |
| GraphWalks | 90% | #5 of 21 | Model card | |
| LongBench v2 | 61% | #11 of 27 | Model card | |
| OpenAI MRCR v2 | 60% | #5 of 9 | Model card | |
| IFBench | 69% | #45 of 70 | Model card | |
| SimpleQA Verified | 31% | #15 of 22 | Model card | |
| MultiChallenge | 53% | #26 of 36 | Model card | |
| LongFact | 98% | #5 of 5 | Model card | |
| CorpusQA | 82% | #1 of 1 | Model card | |
| MAI human side-by-side evaluation | 0.07 | #1 of 1 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AIME 2025 | 97% | #17 of 143 | Model card | |
| GPQA Diamond | 84.2% | #79 of 284 | Model card | |
| AIME 2026 | 94.5% | #15 of 34 | Model card | |
| HMMT 2026 | 84.9% | #12 of 18 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveCodeBench v6 | 87.7% | #12 of 76 | Model card | |
| SWE-bench Verified | 73.5% | #54 of 150 | Model card | |
| SWE-Bench Pro | 52.8% | #56 of 77 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Berkeley Function Calling Leaderboard (BFCL) | 72% | #15 of 47 | Model card | |
| Terminal-Bench 2.0 | 46% | #45 of 61 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| HealthBench Professional | 35% | #25 of 25 | Model card | |
| MedXpertQA | 43% | #2 of 2 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AIR-Bench 2024 | 88% | #1 of 2 | Model card | |
| CyberSecEval Autocomplete | 63% | #2 of 3 | Model card | |
| CyberSecEval Instruct | 63% | #3 of 3 | Model card |
The benchmarks the lab chose to headline, with the scores it gave.
Every result in the longer documents the lab published with it. 9 of them aren't in the launch post.