Methodology
How prices, scores and recommendations are produced.
Where prices come from
Every price is read from the provider's own pricing page, never from a third-party aggregator. The verification date is stored alongside the price, and the source link is kept for each entry.
Prices are US dollars per million tokens. Where a provider uses tiered pricing (a higher rate above some context length), both tiers are stored, and the model page shows them separately.
The "blended" rate used for rankings assumes three input tokens for every output token, which is roughly conversational usage. Your own ratio may differ: use the raw input and output prices if it does.
Why some cells are empty
An empty cell means the data is not published, not that it is zero. We never substitute a missing measurement with zero: that would rank an unmeasured model below a genuinely bad one, which is worse than admitting we do not know.
This matters most for open-weight models, which have no hosted API price at all. Their running cost depends on your hardware, and that is not the same claim as being free.
How scores are computed
Benchmark results come from Artificial Analysis's independent evaluations, run the same way for every model (Intelligence Index, GPQA Diamond, Humanity's Last Exam, Terminal-Bench Hard, SciCode, τ²-Bench, long-context reasoning, instruction following, visual reasoning), with the reasoning setting tested shown on each model page. Speed is their measured median output speed.
The 0-100 scores are derived from those results, normalised against each benchmark's declared scale, then averaged within an axis weighted by how much that benchmark tells you.
Five axes: quality, reasoning, coding, speed and value. Value is quality per dollar, rescaled so the best ratio in the catalogue is 100.
How the wizard ranks
In two stages. First elimination: a model that cannot meet a hard requirement is removed, not down-ranked. Then ranking: the survivors are scored on the axes your task actually depends on.
No AI model is involved in that ranking. It is computed from published data, which makes every recommendation reproducible and checkable.
Weighting by available evidence. A model measured on one axis should not score a perfect result simply because it leads that axis. Scores are therefore shrunk toward a neutral value in proportion to missing data, and every recommendation shows how much of its score was actually backed by measurements.
Limitations
About 90 language models have independent results; image, video and voice models do not yet. A model can be excellent and not yet appear on any leaderboard. Those models are shown separately and flagged as unbenchmarked, rather than hidden.
Prices and capabilities change often. Always confirm against the provider's own page before committing.
Last verified Oct 27, 2026