xbench-ScienceQA: leaderboard

xbench's science question board (xbench.org): graduate-level science questions answered with search turned off and graded by a gemini-3.5-flash judge; the score is the average over five runs. The publisher keeps runs of older model versions and records the reasoning effort of each run; GPT-6 Astra (API modelId openai/gpt-6-astra-pro) and Deepseek-V4-Flash (API release and evaluation dates point to the 0731 checkpoint while modelId and link point to the bare April repo) are held because their checkpoint identities are ambiguous.

Metric: Average score (%). Source: xbench.org. Saturation forecast: Around September 2027. 28 models tracked. Updated 19 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 4.876.4#64
2Claude Opus 5.576.4#9
3Qwen 3.8 Max73.4#66
4Claude Fable 573.4#32
5Claude Opus 4.773#85
6GPT-5.5 ProHigh73#40
7DeepSeek V4.1 FlashMax72.6#69
8Gemini 3.8 FlashHigh72.4#43
9Grok 4.572.4#90
10GPT-5.4 ProxHigh72.2#95
11Gemini 3 Pro (Preview)71.6#120
12Kimi K371.2#61
13Seed 2.1 Pro71#148
14GPT-5.6 Pro Sol71#31
15GPT-5.470.8#138

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=xbench-scienceqa · How It Works · Data refreshed daily, snapshot 2026-11-20.