GPQA Diamond
Graduate-level science reasoning · 16 models · source Epoch AI / Metatext · updated 2026-11
What this benchmark measures
448 multiple-choice questions written and cross-verified by PhDs in physics, chemistry and biology. The answers cannot be found by searching, which makes it the most contamination-resistant knowledge benchmark available.
Why it gets its own table
The leaders are close to saturation (96%+), which reduces discrimination. But it is the key test for whether a small model can replace a large one: the gap between a local 27B and the frontier is clearest here.
Full ranking
| # | Model | Score | Vs. best | Input $ | Output $ | Type |
|---|---|---|---|---|---|---|
| 1 | Gemini 4 ArgonGoogle | 96.4 | Best overall | $2.00 | $10.00 | |
| 2 | GPT-6 AstraOpenAI | 95.8 | 99% | $10.00 | $50.00 | |
| 3 | Claude Sonnet 5.5Anthropic | 95.6 | 99% | $2.00 | $10.00 | |
| 4 | GPT-6.1 SolOpenAI | 95.4 | 99% | $2.00 | $10.00 | |
| 5 | Gemini 3.8 FlashGoogle | 95.4 | 99% | $0.75 | $3.75 | |
| 6 | GPT-6.1 TerraOpenAI | 94.3 | 98% | $1.00 | $5.00 | |
| 7 | Claude Opus 5.5Anthropic | 93.9 | 97% | $4.00 | $20.00 | |
| 8 | Kimi K3 Open weightsMoonshot AI | 93.1 | 97% | $2.00 | $15.00 | Open weights |
| 9 | Grok 4.7xAI | 92.7 | 96% | $2.00 | $6.00 | |
| 10 | Qwen3.8 MaxAlibaba Qwen | 92.7 | 96% | $2.00 | $8.00 | |
| 11 | GLM-5.3 Open weightsZ.ai | 90.9 | 94% | $0.60 | $2.20 | Open weights |
| 12 | MiniMax M3 Open weightsMiniMax | 90.9 | 94% | $0.25 | $1.00 | Open weights |
| 13 | DeepSeek V4.1 Flash Open weightsDeepSeek | 89.8 | 93% | $0.30 | $1.20 | Open weights |
| 14 | Gemma 4 31B Open weightsGoogle | 75.8 | 79% | $0.10 | $0.40 | Open weights |
| 15 | gpt-oss-120b Open weightsOpenAI | 75.8 | 79% | $0.05 | $0.20 | Open weights |
| 16 | MiMo-V2.6 Pro Open weightsXiaomi | 59.5 | 62% | $0.35 | $1.40 | Open weights |