ARENA + VALS + ARTIFICIAL ANALYSIS

Find the right model for your work.

Choose the work. Weigh the evidence. See what the quality costs.

A separate, experimental view

What are you doing?

What matters to you

Weights normalize to 100% · change any weight

Compare Arena evidence Order by Quality margin points below best Min. measured coverage percent of weighted evidence

Estimate a request

Token budget includes any billed reasoning · USD

Input tokens Output tokens Cached input percent of input Max. cost / request optional USD limit Max. first answer optional median seconds

Include predicted benchmark scores Open weights onlyRanking uses recorded benchmark results.

Eligible—

Best composite—

Lowest cost near best—

Fastest first answer near best—

Your ranking

Loading benchmark snapshots…

Show insufficient evidence All benchmark inputs

Rank	Model & settings	Task score	Evidence	Estimated cost	First answer
Reading three sources of evidence…

A score you can trace.

Tasks set the weights. Each task contains fixed benchmark slots. Missing slots contribute a neutral 50. Recorded coverage counts the weights of measured slots. A model needs at least two measured benchmark families, plus your minimum coverage, to receive a rank. Search and creator filters never change scores or the reference cohort.

Repeated evidence gets one slot. AA and Vals Terminal-Bench results share a benchmark slot. Each source is summarized before sources are combined. Release view takes the median across settings; configuration view only uses that setting. Vals effort settings absent from the reference stay separate.

Predictions are optional. Measured results take priority. An estimate uses half its slot weight, with the rest neutral. Multiple estimates from one underlying test share a capped evidence weight. Estimates never increase measured coverage. These settings are explicit heuristics, not validated confidence probabilities.

Cost and speed describe a workload. Cost uses your input, output and cache assumptions, with OpenRouter rates when matched, otherwise AA rates. Request fees and published long-context tiers are included; external tool and media costs are excluded. AA first-answer medians describe its test workload, not a response-time guarantee. Vals task duration is listed separately.

Each benchmark is normalized against a fixed cohort with one median observation per release. Different benchmarks have different cohorts; a percentile is a relative comparison. Release summaries mix harnesses and settings and cannot predict an exact deployment. Composite scores, weight choices and quality margins are exploratory; available score ranges are not statistical confidence intervals.

Snapshots & provenance

The Arena dataset and OpenRouter catalog refresh with the daily site build. Vals and AA remain the supplied snapshots until explicitly reimported. Benchmark links use the benchmark definitions; the supplied AA cache points every result at one generic page.