Put models to the test.
A clearer picture of what your model can do. Run a quick benchmark and turn the results into insight.
Set up your run
Choose a model, benchmark and difficulty.
Reuse the same seed, type and run length to compare identical Advanced questions.
✳Runs may use credits. Actual cost depends on the model and tokens used.
Your report card
A little clarity goes a long way.
Performance breakdown
Open a category to inspect its checks and response.
Your report starts here
Choose a model and run the benchmark to see its results.
Good to know
A useful snapshot, not the whole story.
These are custom English-language tests, not industry-standard benchmarks. A perfect score is possible and only means the model passed these checks. Scores don’t measure safety, creativity or overall real-world performance.
See how scoring works↓Each request contains three or five equally weighted checks. Requests cycle through selected categories; counts that aren’t multiples of the number of categories give earlier categories an extra request. Bars combine all requests for each category. The overall score weights each check equally, so categories with more checks count more. Advanced questions use generated values and computed answer keys, but share templates. Basic and intermediate questions repeat on longer runs. Non-format checks allow a JSON object surrounded by extra text; the formatting test requires valid JSON only, with the requested values and types. Failed or empty requests stop the run instead of counting as wrong. Grades: A ≥ 90, B ≥ 80, C ≥ 70, D ≥ 60, otherwise F. Compare the same suite version, type, level and run length across models, and repeat runs. Generated data should be reviewed before training; keep evaluation data separate from training data.