Benchmark leaders by category
The top model on each benchmark in the index, grouped by benchmark category.
What the data shows
Different models lead different benchmarks. That is the main argument against choosing a single model for every task: the leader on coding is not necessarily the leader on knowledge or preference.
The table is a map of strengths, not a single ranking.
How to read it
- Match the benchmark category to your task.
- Check the leader's price and context.
- Compare the top three, not just the winner.
- Test the leader on your own data.
Benchmark leaders
| Benchmark | Leader | Score | Category |
|---|---|---|---|
| MMLU-Pro | GPT-5.2 Pro | 91.6 | Knowledge |
| GPQA Diamond | GPT-5.2 Reasoning | 87.0 | Reasoning |
| SWE-bench Verified | GPT-5.2 Pro | 81.7 | Coding |
| HumanEval | GPT-5.2 Pro | 96.8 | Coding |
| MATH-500 | GPT-5.2 Max | 96.7 | Mathematics |
| AIME 2025 | Claude Opus 4.5 Pro | 91.4 | Mathematics |
| LiveCodeBench | Claude Sonnet 4.5 Reasoning | 85.1 | Coding |
| IFEval | Claude Opus 4.5 Pro | 96.1 | Instruction following |
| TruthfulQA | GPT-5.1 Max | 91.1 | Safety |
| HellaSwag | GPT-5.2 Max | 96.1 | Commonsense |
| MMMU | GPT-5.2 Max | 86.9 | Multimodal |
| Arena Elo | Claude Opus 4.5 Max | 1501.9 | Preference |
Methodology
This story uses the AI & LLM Index dataset: 600 models, 40 providers and 12 benchmarks. Prices are listed per million tokens. Scores are modelled estimates on a 0–100 scale and are not official benchmark results. Figures are rounded for readability.
Related
FAQ
What does the Benchmark leaders by category data show?
The top model on each benchmark in the index, grouped by benchmark category.
Are the numbers official?
No. Prices and scores are compiled for this demo index. Benchmark values are modelled estimates, not official results. Verify figures with the provider or benchmark owner before making decisions.
How should I use this ranking?
Use it to form a shortlist. Check the specific model pages, compare price and context, and test the shortlist on your own examples before committing.