Skip to content

Benchmark

CritPT

Science benchmarks measure how well models answer science questions. Scores are percentages. Higher is better.

Models measured
156
Scores recorded
268
Sources
3
IQ weight
2.67

Score distribution

0.98%
Mean
0.106%
Median
4%
Std dev
0%
10th percentile
0.314%
90th percentile
0%
Min
0%31.4%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 2.67 for chat models, measured over 156 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
101DeepSeek V4 Flash 0731max· Epoch AI Benchmarking Hub data16.6%
102GPT-5.6 Lunahigh· Epoch AI Benchmarking Hub data16.6%
103Beam 1.0 501B A23B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim16.3%
104GPT-6 Sollow· Epoch AI Benchmarking Hub data16.3%
105Claude Sonnet 5xhigh· Epoch AI Benchmarking Hub data15.4%
106GLM 5.3 Flash· Epoch AI Benchmarking Hub data15.4%
107GPT-6 Lunahigh· Epoch AI Benchmarking Hub data15.4%
108Grok 4.5high· Epoch AI Benchmarking Hub data15.4%
109GPT-6.1 Terralow· Epoch AI Benchmarking Hub data15.1%
110Muse Spark 1.1xhigh· Epoch AI Benchmarking Hub data15.1%
111Muse Spark 1.1· Epoch AI Benchmarking Hub data15.1%
112GPT-5.6 Sollow· Epoch AI Benchmarking Hub data14.9%
113Muse Spark Open 816B A42Bxhigh· Epoch AI Benchmarking Hub data14.9%
114GLM 5.3low· Epoch AI Benchmarking Hub data14.6%
115DeepSeek V4.1 Flashmax· Epoch AI Benchmarking Hub data14.3%
116Gemini 3.7 Flashhigh· Epoch AI Benchmarking Hub data14.3%
117Grok 4.7low· Epoch AI Benchmarking Hub data13.4%
118Qwen3.7 Max· Epoch AI Benchmarking Hub data
Values from 2 sources
13.4%
119Gemini 3.5 Flashhigh· Epoch AI Benchmarking Hub data13.1%
120Claude Haiku 5.5medium· Epoch AI Benchmarking Hub data12.9%
121DeepSeek V4 Promax· Epoch AI Benchmarking Hub data12.9%
122GPT 5high· Epoch AI Benchmarking Hub data12.6%
123Gemini 3.8 Flashmedium· Epoch AI Benchmarking Hub data12.3%
124Claude Opus 4.7max· Epoch AI Benchmarking Hub data12%
125MiMo V2.6 Flash· Epoch AI Benchmarking Hub data12%
126Claude Sonnet 5.5low· Epoch AI Benchmarking Hub data11.4%
127Qwen3.7 MaxQwen3.7-Max release · recommended reasoning-effort=xhigh system prompt· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.11.4%
128Muse Spark· Epoch AI Benchmarking Hub data11.3%
129Gemini 3.6 Flashhigh· Epoch AI Benchmarking Hub data10.6%
130GPT-6 Lunamedium· Epoch AI Benchmarking Hub data10.6%
131Mistral Large 4· Epoch AI Benchmarking Hub data10.6%
132DeepSeek V4 Prohigh· Epoch AI Benchmarking Hub data10%
133GPT-5.4 minixhigh· Epoch AI Benchmarking Hub data10%
134Kimi K2.7 Code· Epoch AI Benchmarking Hub data10%
135Gemini 3.7 Flashmedium· Epoch AI Benchmarking Hub data9.4%
136GPT-5.6 Terralow· Epoch AI Benchmarking Hub data9.4%
137GPT 5.4 Nanoxhigh· Epoch AI Benchmarking Hub data9.3%
138Claude Haiku 5.5low· Epoch AI Benchmarking Hub data9.1%
139Grok Build 0.1· Epoch AI Benchmarking Hub data9.1%
140Qwen3.7 Plus· Epoch AI Benchmarking Hub data
Values from 2 sources
9.1%
141Claude Sonnet 5medium· Epoch AI Benchmarking Hub data8.6%
142Inkling Small· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim
Values from 2 sources
8.3%
143GPT 5.5low· Epoch AI Benchmarking Hub data8%
144Grok 4.3high· Epoch AI Benchmarking Hub data8%
145Kimi K2.6· Epoch AI Benchmarking Hub data8%
146DeepSeek V4 Flashmax· Epoch AI Benchmarking Hub data7.1%
147Gemini 3 Pro Preview· Epoch AI Benchmarking Hub data6.9%
148Qwen3.7 PlusQwen3.7-Plus release · recommended reasoning-effort=xhigh system prompt· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.6%
149Gemini 3.7 Flashlow· Epoch AI Benchmarking Hub data5.7%
150Grok 4.6low· Epoch AI Benchmarking Hub data5.7%

Showing 101–150 of 268 results.

Data credits: Data by LLM Stats · Epoch AI Benchmarking Hub, CC BY 4.0 (licence) · BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.