Public third-party benchmarks in one place, comparing leading AI models on intelligence, cost and speed with the same yardstick.
Scores from Artificial Analysis Intelligence Index v5.0; click a column header to sort.
| Rank | Model / configuration |
|---|
Ties use the rounded values published by the source; a 1–2 point gap is not a reliable difference. Ranks apply only within the current filters.
Released models without an independent score, special product modes, or invitation-only models. No score doesn’t mean a weak model.
This site collects public third-party AI model benchmarks and compares models on intelligence, cost and speed with the same yardstick, keeping versions and reasoning effort separate. We don’t run benchmarks ourselves.
Data updated 2026-11-03; updated weekly. Each metric’s check date is shown in the model details.
Intelligence: Artificial Analysis Intelligence Index v5.0, built from 9 evaluations (knowledge work, coding, science and general, 25% each); ties share a rank.
Cost: only configurations meeting a score bar (35 by default), sorted by AA’s average cost per evaluation task, low to high.
Speed: output speed measured by AA (~10k input, 72-hour median), or time to first token. AA methodology ↗
The score is not an IQ or accuracy rate, and the benchmarks lean toward English. Cost is what AA spent running the benchmark, not your per-prompt price. Speed is measured on US servers and may differ where you are; the first token may be thinking output rather than the answer.
Vendor-reported results appear only in details and are not ranked. Legacy v4.3.2 scores are shown separately and never mixed with v5. Arena human-preference scores are shown separately and not merged into the score.
This site is not affiliated with Artificial Analysis, Arena or any model vendor. Data and trademarks belong to their respective owners; sources are linked below. Data may contain errors or be out of date. For reference only; not purchasing advice.
| Arena rank | Entry | Preference score | Rank range |
|---|
Test candidate models on your own tasks with the same inputs and acceptance criteria, and count the cost of failures and retries.
Real cost per task = total spend ÷ tasks that finally passed.
Calculated only in your browser; nothing is sent or stored.
Every number traces back to the public pages below; check dates are listed with each entry.