Compare top models across the benchmark suite that best represents reasoning performance. Use this page as the fastest way to inspect the relevant tests, then jump into the full matrix when you want broader context.
4
Benchmarks in category
44
Models with coverage
2
Benchmarks with human baseline
2
Saturated benchmarks
The current benchmark set in this category, with context on what each test captures.
One of the best discriminators between models. Scores range widely (40-85%), making it highly informative for comparing reasoning ability.
Tests commonsense scientific reasoning. Largely saturated for frontier models but still useful for comparing mid-tier and open-source models.
Fundamental commonsense reasoning test. Saturated for frontier models (>95%) but useful for evaluating smaller models.
The hardest academic benchmark — top models still fail 60-65% of questions. Shows how far we are from genuine expert-level reasoning.
Expert-level science reasoning across biology, chemistry, and physics at PhD level. Questions are designed to be 'Google-proof' — even domain experts with web access struggle.
Why it matters
One of the best discriminators between models. Scores range widely (40-85%), making it highly informative for comparing reasoning ability.
| # | Model | Score |
|---|---|---|
| 1 | 🥇GPT-5.4 | 88.5% |
| 2 | 🥈Claude Opus 4.6 | 88.4% |
| 3 | 🥉o3 | 87.7% |
| 4 | GPT-5.2 | 87.0% |
| 5 | Claude Opus 4.5 | 86.2% |
| 6 | Gemini 3 Pro | 86.0% |
| 7 | GPT-5 | 85.0% |
| 8 | Claude 3.7 Sonnet | 84.8% |
| 9 | Gemini 2.5 Pro | 84.0% |
| 10 | Claude Opus 4 | 82.6% |
| 11 | Claude Sonnet 4.6 | 82.0% |
| 12 | Grok 4 | 82.0% |
| 13 | o4-mini | 81.4% |
| 14 | o3-mini | 79.7% |
| 15 | Claude Sonnet 4.5 | 78.5% |
| 16 | o1 | 78.0% |
| 17 | DeepSeek R1-0528 | 76.0% |
| 18 | Claude Sonnet 4 | 74.2% |
| 19 | DeepSeek R1 | 71.5% |
| 20 | Gemini 3 Flash | 70.0% |
| 21 | Grok 3 | 68.2% |
| 22 | Gemini 2.5 Flash | 65.3% |
| 23 | Claude 3.5 Sonnet | 65.0% |
| 24 | DeepSeek V3 (March 2025) | 62.5% |
| 25 | Gemini 2.0 Flash | 62.1% |
| 26 | Llama 4 Maverick | 61.0% |
| 27 | o1-mini | 60.0% |
| 28 | Gemini 1.5 Pro | 59.1% |
| 29 | DeepSeek V3 | 59.1% |
| 30 | Phi-4 | 56.1% |
| 31 | GPT-4.1 | 56.0% |
| 32 | GPT-4o | 53.6% |
| 33 | Mistral Large 2 | 52.5% |
| 34 | Claude Haiku 4.5 | 52.0% |
| 35 | Llama 3.1 405B | 51.1% |
| 36 | Llama 3.3 70B | 50.5% |
| 37 | Claude 3 Opus | 50.4% |
| 38 | Qwen 2.5 72B | 49.0% |
| 39 | GPT-4 Turbo | 48.0% |
| 40 | Llama 3.1 70B | 46.7% |
| 41 | Claude 3.5 Haiku | 41.6% |
| 42 | GPT-4o mini | 40.2% |
Performance Tiers
Model Types
Saturated benchmarks have top models clustered above 90%, making them less useful for comparison.
Scores sourced from official model cards, technical reports, and third-party evaluations (Artificial Analysis, LMSYS Arena). Last updated: 2026-03-07. Some scores are approximate.
AI benchmarks are grouped into categories like coding, math, reasoning, knowledge, and safety. Each category contains multiple standardized tests that measure specific aspects of model performance. This page focuses on one category so you can compare models within a specific skill area.
Each benchmark has its own scoring method - accuracy percentage, pass rate, Elo rating, or normalized score. We display raw scores from official evaluations and community-run tests. Scores are updated hourly as new evaluation results become available.
A saturated benchmark is one where top models score near the maximum (typically above 95%). This means the benchmark no longer effectively differentiates between the best models, and newer, harder benchmarks are needed to measure progress.