What's new
now jobs Jobs market snapshot refreshed now data Recomputed benchmark-weighted quality scores now data Updated speed measurements now data Synced Chatbot Arena benchmark track now data Validated official pricing snapshots now data Pulled latest OpenRouter price index 25 May digest OpenGuardrails: An Open-Source Context-Aware AI Guardrails Platform 25 May digest Published the 2026-05-25 daily digest 25 May digest Workbench Launches Open Source BullMQ Dashboard For Node Backends 24 May digest SpecBench Tests Reward Hacking In Long Horizon Coding Agents now jobs Jobs market snapshot refreshed now data Recomputed benchmark-weighted quality scores now data Updated speed measurements now data Synced Chatbot Arena benchmark track now data Validated official pricing snapshots now data Pulled latest OpenRouter price index 25 May digest OpenGuardrails: An Open-Source Context-Aware AI Guardrails Platform 25 May digest Published the 2026-05-25 daily digest 25 May digest Workbench Launches Open Source BullMQ Dashboard For Node Backends 24 May digest SpecBench Tests Reward Hacking In Long Horizon Coding Agents

GPQA Diamond

reasoning

GPQA Diamond (Graduate-Level Google-Proof Q&A) contains extremely difficult science questions that even domain experts find challenging. "Diamond" refers to the hardest subset where expert validators achieved less than 65% accuracy.

View paper / source

38

Models Tested

91.0

Best Score

69.0

Average Score

0–100

Scale Range

1.5x

Weight

How It Works

Models answer open-ended graduate-level questions in physics, chemistry, and biology. Questions are designed to be "Google-proof" — they cannot be answered by simply searching the internet. Expert validators with PhDs in the relevant field verify question difficulty.

Why It Matters

GPQA Diamond is one of the most reliable indicators of deep scientific reasoning. Because even domain experts struggle with these questions, high scores genuinely indicate advanced reasoning capabilities rather than memorisation.

Limitations

Small dataset size (~200 questions) means scores can be noisy. Heavy STEM focus doesn't capture humanities or creative reasoning. Expert disagreement on some answers.

Leaderboard — GPQA Diamond

# Model Provider Score
🥇 GPT-5.2 Pro OpenAI 91.0
🥈 GPT-5.2 OpenAI 89.0
🥉 GPT-5 Pro OpenAI 88.0
4 o3 Pro OpenAI 87.5
5 GPT-5 OpenAI 86.0
6 o3 OpenAI 83.3
7 Grok 4 xAI 82.0
8 o4 Mini OpenAI 81.4
9 Gemini 3 Pro Google 80.0
10 o1 OpenAI 78.0
11 Claude Opus 4.5 Anthropic 78.0
12 R1 0528 DeepSeek 76.0
13 o3 Mini OpenAI 75.0
14 Claude Opus 4 Anthropic 72.1
15 Claude Sonnet 4.5 Anthropic 72.0
16 Qwen3 Max Alibaba 72.0
17 R1 DeepSeek 71.5
18 Gemini 2.5 Pro Preview 06-05 Google 68.4
19 Grok 3 xAI 68.2
20 Claude 3.7 Sonnet Anthropic 68.0
21 Claude Sonnet 4 Anthropic 67.5
22 GPT-4.1 OpenAI 66.3
23 Phi-4 Reasoning Microsoft 65.8
24 Claude 3.5 Sonnet Anthropic 65.0
25 QwQ 32B Alibaba 63.0
26 GPT-4.5 OpenAI 62.0
27 DeepSeek V3 0324 DeepSeek 62.0
28 o1-mini OpenAI 60.0
29 Gemini 2.5 Flash Google 59.2
30 DeepSeek V3 DeepSeek 59.1
31 Mistral Large Mistral 58.0
32 Phi 4 Microsoft 56.1
33 Llama 4 Maverick Meta 56.0
34 GPT-4o (2024-05-13) OpenAI 53.6
35 Llama 3.1 405B Meta 51.1
36 Llama 3.3 70B Instruct Meta 50.7
37 Claude 3 Opus Anthropic 50.4
38 Qwen2.5 72B Instruct Alibaba 49.0
All Benchmarks