GPQA Diamond (Vals protocol average)
General QA · 2026-07-10
Vals AI's GPQA Diamond Overall panel averages each model's Few-Shot CoT and Zero-Shot CoT accuracy. It is separated from either individual prompting protocol so the derived average does not create conflicting observations on the canonical GPQA Diamond panel.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.1 Pro Preview | 95.5 |
| GPT-5.6 Sol | 95.2 |
| Grok 4.6 | 94.7 |
| Gemini 3.8 Flash | 94.4 |
| Gemini 3.7 Flash | 93.9 |
| Qwen3.8-Max | 93.7 |
| Claude Fable 5.1 | 93.4 |
| Claude Opus 5 | 93.4 |
| Gemini 3.6 Flash | 93.4 |
| GPT-5.5 | 93.2 |
| Grok 4.5 | 92.9 |
| Kimi K3 | 92.9 |
| Gemini 3.5 Flash | 92.7 |
| MiniMax M3 | 92.7 |
| Opus 4.8 | 92.4 |
| DeepSeek V4 Pro 0813 | 92.4 |
| Gemini 3 Pro Preview | 91.7 |
| GPT-5.2 | 91.7 |
| GPT-5.4 | 91.7 |
| GPT-5.6 Luna | 91.7 |
| Grok 4.3 | 91.4 |
| Muse Spark 1.1 | 91.2 |
| GPT-5.6 Terra | 90.9 |
| Opus 4.7 | 90.2 |
| Qwen3.7-Max (source unspecified) | 90.2 |
| DeepSeek-V4-Flash-0731 | 89.9 |
| Opus 4.6 | 89.6 |
| Muse Spark | 89.6 |
| DeepSeek-V4-Pro | 89.4 |
| Kimi K2.6 | 89.1 |
| Sonnet 5 | 88.9 |
| Qwen3.8 27B | 88.9 |
| Grok 4.20 | 88.6 |
| GLM-5.3 | 88.1 |
| Grok 4 | 88.1 |
| Gemini 3 Flash Preview | 87.9 |
| Qwen3.5 Plus (2026-02-15) | 87.4 |
| Qwen3.6 Plus (2026-04-02) | 87.4 |
| Inkling | 87.1 |
| GPT-5.1 | 86.6 |