How we measure AI.

LLM Stats combines independently evaluated results with clearly labelled external evidence. We prioritize consistent comparisons, source provenance, and version history so readers can understand what each result represents.

Evidence is reviewed before publication and may be updated when a model, benchmark, or source materially changes.

Arenas capture human preferences through blind comparisons. The experience is designed to reduce the influence of model identity and presentation order while supporting a broad range of tasks and modalities.

Quality controls help identify unreliable activity and preserve the usefulness of aggregate preference signals.

Relative performance is combined with an uncertainty-aware statistical approach. Stronger and more consistent evidence has more influence, while limited or conflicting evidence is treated more cautiously.

Missing observations are not treated as failures. Ratings are intended for comparison within their stated context, not as a universal measure of model quality.

The catalogue covers major capability areas including coding, reasoning, multimodal understanding, long-context work, and professional domains. Coverage evolves as evaluation practices and available models change.

Results from different domains remain distinct where combining them would hide meaningful differences in model behavior.

Benchmarks and preference data are imperfect proxies for real deployments. Results can be affected by task selection, model updates, provider behavior, evaluation design, and the availability of comparable evidence.

Rankings should be considered alongside cost, latency, safety, reliability, and requirements specific to the intended use case.

We keep implementation parameters, operational controls, and other sensitive methodology details out of this public overview. Researchers, partners, and customers can contact us for further details.