36%
best score so far on CritPt (physics research), the hardest of these 15 reasoning benchmarks for AI
What the chart shows
Each row is one reasoning benchmark in Epoch AI’s hub. The best score so far is highest on ARC-AGI-1 (98%, Claude Fable 5 (xhigh), 2026) and lowest on CritPt (physics research) (36%, Gemini 4 Carbon (high), 2026).
ARC-AGI-1, HellaSwag, ARC (AI2) Challenge and ARC-AGI-2 are at 90% or more, close to their ceiling, so they no longer separate the leading models. Newer benchmarks are built to stay hard, which is why their scores start low.
Benchmarks with their own chart pages: ARC-AGI-2.
Key numbers
| Measure | Value |
|---|---|
| ARC-AGI-1 (2026) | 98% |
| HellaSwag (2023) | 95% |
| ARC (AI2) Challenge (2024) | 95% |
| ARC-AGI-2 (2026) | 95% |
| WinoGrande (2024) | 89% |
| BIG-Bench Hard (2024) | 89% |
| PIQA (2024) | 89% |
| Mystery game puzzles (2026) | 84% |
| LiveBench (2025) | 82% |
| SimpleBench (2026) | 82% |
| Chess puzzles (2026) | 76% |
| Adversarial NLI (2023) | 58% |
| CommonsenseQA 2 (2023) | 57% |
| EnigmaEval (2026) | 39% |
| CritPt (physics research) (2026) | 36% |
How to read it
- Each cell is the best score among models released that year, from Epoch AI’s own runs or the results it collects; settings differ by benchmark.
- Only benchmarks scored as a share of questions or tasks, with results from at least two release years, are included.
- Choose a benchmark from the list above the chart; the full grid is in the downloadable image and CSV.
Source: Epoch AI, 2026: Benchmarking Hub, read 3 November 2026. Chart and text by AI Adoption.
Numbers checked against the published data on 3 November 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 3 November 2026.
Download the data (CSV) · Cite this chart
Discussion