AI Benchmarks
Plain-English explainers for every AI benchmark we cite, the caveats that come with them, and a side-by-side dashboard of published scores across 0 models.
Why benchmarks disagree (and what to do about it)
No single benchmark settles which AI model is best. Benchmarks measure narrow tasks under specific conditions. Different ones reward different things — knowledge, reasoning, code, style, latency, cost — and a model that wins on one usually loses on another.
Read our Benchmark Methodology for the caveats we apply, and never treat a leaderboard ranking as a verdict.
Benchmark explainers
Knowledge
Math
Coding
164 Python coding problems with hidden unit tests.
Continuously-updated competitive-programming problems from LeetCode, AtCoder, and Codeforces.
Real GitHub issues from 12 popular Python repos that the model must resolve end-to-end.
500-instance cleaned subset of SWE-Bench with verified test reliability.
Agents
Human Preference
Multimodal
Model benchmark dashboard
Side-by-side benchmark scores reported by model developers. Sort or filter to find top performers. Every score is published on the source model's documentation; see each model page for the citation.
0 models with benchmark data
Frequently Asked Questions
What benchmarks are shown in this dashboard?
We display publicly reported benchmarks including MMLU, HumanEval, GSM8K, MATH, HellaSwag, and other standard AI evaluation metrics. Not all models report all benchmarks.
How are benchmark scores collected?
Benchmark data comes from official model releases, research papers, and third-party evaluation platforms. Scores are updated as new data becomes available.
Can I compare benchmarks across different model categories?
Yes. Use the category filter to compare LLMs, code models, or multimodal models. You can also sort by any benchmark column to find top performers.