Enterprise AI in Numbers · Capability and cost
Top score on Humanity's Last Exam, the benchmark of expert-level questions
54.8%
54.8% (GPT-6 Astra) as of September 3, 2026, according to Epoch AI, Benchmarking hub, up from 46.5% on September 1, 2026. The highest accuracy any evaluated model reaches on Humanity's Last Exam, questions written by subject experts to sit beyond what earlier models could answer, with its lab.
What it means for leaders
Expert-level accuracy is rising fast but remains well short of reliable. Use these models to draft and check expert analysis, with the expert still accountable for the answer.
What would change the reading: The pace of gains, and whether calibration (knowing when it is wrong) improves with accuracy.
The trend
Every reading
| Date | Value | Change | Detail | Source |
|---|---|---|---|---|
| 54.8% | +8.3 points | GPT-6 Astra | Source | |
| 46.5% | +0.1 points | Claude Fable 5.1 | Source | |
| 46.4% | +8.9 points | Gemini 3.1 Pro | Source | |
| 37.5% | +5.9 points | Gemini 3 Pro | Source | |
| 31.6% | +6.3 points | GPT-5 Pro | Source | |
| 25.3% | +3.7 points | GPT-5 | Source | |
| 21.6% | +1.3 points | Gemini 2.5 Pro (Jun 2025) | Source | |
| 20.3% | +2.1 points | o3 | Source | |
| 18.2% | +10.1 points | Gemini 2.5 Pro (Mar 2025) | Source | |
| 8.1% | +0.1 points | o1-pro | Source | |
| 8.0% | +3.4 points | o1 | Source | |
| 4.6% | — | Gemini 1.5 Pro (Sept 2024) | Source |
Definition and method
- What it measures
- The highest accuracy any evaluated model reaches on Humanity's Last Exam, questions written by subject experts to sit beyond what earlier models could answer, with its lab.
- Population
- AI models evaluated on Humanity's Last Exam, as compiled by Epoch AI
- Updated
- When new models are evaluated
- Source
- Epoch AI, Benchmarking hub · CC BY 4.0
Caveats
- Scores come from independent evaluators compiled by Epoch AI, not from the labs, and depend on the settings and tools each run used.
- The leader is the best result so far, dated by the release of the model that set it; a model evaluated late can take the lead after its release date.
- Results are compiled from several evaluators, so settings vary between models.
Cite this
Enterprise AI in Numbers, "Top score on Humanity's Last Exam, the benchmark of expert-level questions", https://lucianoayres.com/ai-brief/data/expert-reasoning-benchmark-leader/, retrieved October 25, 2026. Data: Epoch AI, Benchmarking hub.