GPQA Diamond
GPQA Diamond is the "graduate-level google-proof" science benchmark — 198 questions written by domain PhDs to resist web search. Top frontier models are now in the 85–90% range, comfortably above the ~65% human-expert baseline. The benchmark that gave first hard evidence of super-human performance on specialized scientific reasoning.
Category: reasoning · Source: github.com ↗
| # | Model | Score | Setting | Measured | Source |
|---|---|---|---|---|---|
| 01 |  Gemini 3.5 | 91.2% | 0-shot CoT, Deep Think | May 20, 2026 | link ↗ |
| 02 |  Claude 4.7 OpusAnthropic | 90.1% | 0-shot CoT | Apr 16, 2026 | link ↗ |
| 03 |  GPT-5.5OpenAI | 89.4% | 0-shot CoT | Apr 23, 2026 | link ↗ |
| 04 | Gemini 3 Pro | 88.6% | 0-shot CoT | Apr 22, 2026 | link ↗ |
| 05 | o3 OpenAI | 87.7% | 0-shot CoT | Apr 16, 2025 | link ↗ |
| 06 |  Grok 4xAI | 87.5% | 0-shot CoT | Jul 9, 2025 | link ↗ |
| 07 | GPT-5 OpenAI | 87.3% | 0-shot CoT | Aug 7, 2025 | link ↗ |
| 08 | Claude 4.6 Sonnet Anthropic | 85.8% | 0-shot CoT | Feb 17, 2026 | link ↗ |
| 09 | Gemini 2.5 Pro | 84.0% | 0-shot CoT | Jun 17, 2025 | link ↗ |
| 10 |  Qwen3.7-MaxAlibaba | 83.0% | 0-shot CoT | May 20, 2026 | link ↗ |
| 11 |  DeepSeek-V4-ProDeepSeek | 82.4% | 0-shot CoT | Apr 22, 2026 | link ↗ |
Other model leaderboards
- gpt.buzz Composite Model Index — cross-benchmark ranking
- MMLU-Pro — Multitask language understanding across 14 disciplines — the harder, contamination-resistant successor to MMLU. 10-option multiple choice with stronger distractors.
- HumanEval — OpenAI's 164 Python programming problems with hidden unit tests — the classic code-generation pass@1 benchmark.
- Aider Polyglot — Multi-language code edit benchmark — solve real Exercism problems across Python, JavaScript, Rust, Go, Java, C++ via edit-and-test loops.
- AIME 2025 — American Invitational Mathematics Examination — competition math problems requiring multi-step reasoning. Reasoning-tier models score in the 90s; non-reasoning in the 50s.
- LiveCodeBench — Contamination-free code benchmark — only problems published AFTER a model's training cutoff. Refreshed monthly.
- MMMU — Massive Multi-discipline Multimodal Understanding — 11.5k college-level multimodal questions across 30 subjects (art, business, science, medicine, etc.).