Rankings
Compare AI model capability and cost, or switch to the open-source projects gaining momentum now.
Defaults to the top 30 by capability; add or remove any number of models.
Price blends input and output at 1:3, approximating conversational usage.
Open license means a reported MIT, Apache or similar license; it does not establish that training data and code are open. Open weights includes licenses with additional restrictions. Missing or conflicting licenses remain unverified.
Include models with limited overall evidenceWorldAttention combines relative performance across five public evaluation sources into a capability score from 0 to 100. Higher scores indicate stronger overall performance.
We compare models within each evaluation source, combine those comparisons using the source weights below, and calibrate scores against a fixed group of reference models.
Data completeness: The combined weight of evaluation sources available for a model. For example, coverage from Artificial Analysis, LiveBench and LMArena gives a data completeness of 70%.
The default view includes models with at least three evaluation sources. Enable “Show models with insufficient evidence” to include models with one or two sources and their relative capability estimates.
Ordered by composite model score, highest first.
The composite score reflects relative capability across public evaluations. Higher scores indicate stronger overall performance. “How model scores work” above explains the method and evaluation sources.
Kimi · 2026-04-20
MiniMax · 2026-06-01