HELM Capabilities - GPQA: leaderboard
HELM Capabilities run of GPQA, graduate-level science multiple-choice questions scored by chain-of-thought accuracy.
Metric: COT correct. Source: crfm.stanford.edu. Saturation forecast: Around January 2027. 68 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 80.27 | #66 |
| 2 | GPT-5 | 79.15 | #98 |
| 3 | GPT-5 Mini (2025-08-07) | 75.56 | #172 |
| 4 | O3 (2025-04-16) | 75.34 | #121 |
| 5 | Gemini 2.5 Pro (Preview 03-25) | 74.89 | #123 |
| 6 | O4 Mini (2025-04-16) | 73.54 | #179 |
| 7 | Grok 4 | 72.65 | #181 |
| 8 | Claude Opus 4 (20250514) (Thinking) | 70.85 | #154 (Claude Opus 4 (20250514)) |
| 9 | Claude Sonnet 4 (20250514) (Thinking) | 70.63 | #213 (Claude Sonnet 4 (20250514)) |
| 10 | Claude Sonnet 4.5 | 68.61 | #142 |
| 11 | GPT-OSS-120B | 68.39 | #341 |
| 12 | GPT-5 Nano | 67.94 | #425 |
| 13 | Grok 3 Mini Beta | 67.49 | #296 |
| 14 | DeepSeek R1 0528 | 66.59 | #222 |
| 15 | Claude Opus 4 (20250514) | 66.59 | #154 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-capabilities-gpqa · How It Works · Data refreshed daily, snapshot 2026-10-20.