Keep settings attached
“High,” “max,” tool access and fallback runs can change both results and cost. They are part of the tested configuration.
AI BENCHMARKS
Reasoning, writing, building websites and generating media are different jobs. Here’s how the leaders compare in each.
Results from Artificial Analysis and Arena. Artificial Analysis refreshes daily only while its methodology matches the reviewed v5.0.1 reference. Arena scores are reviewed in a browser. Check source dates, model settings and uncertainty when comparing scores.
Artificial Analysis
A composite of knowledge work, coding, science and general evaluations. Higher scores are better; these are index points, not percent correct.
Intelligence Index v5.0.1
Effort settings and fallback runs affect results. An aggregate score can hide weaknesses on your particular task. Version 5.0 changed the evaluation suite and weighting (5.0.1 only regraded one test); compare only scores from the same major version.
Full results at Artificial Analysis ↗Arena
People compare anonymous model responses and choose the answer they prefer.
Overall · Style Control · All licenses
Source date shown by Arena: 2026-11-09. Checked dates record our review, not a new source release.
Preference is not a factual-accuracy score. A listing does not guarantee access: some entries are previews, and others need a paid plan. Uncertainty ranges overlap for many entries; their order is not proof of a meaningful difference.
Full results at Arena ↗Arena
Head-to-head preferences for generated web applications.
WebDev Overall · All licenses
Source date shown by Arena: 2026-11-10. Checked dates record our review, not a new source release.
This measures web-building output, not every kind of coding, security or maintenance work. Uncertainty ranges overlap for many entries; their order is not proof of a meaningful difference.
Full results at Arena ↗Arena
Human preferences between images generated from prompts.
Text-to-image Overall · All licenses
Source date shown by Arena: 2026-11-07. Checked dates record our review, not a new source release.
A high rating does not guarantee accurate text, product details or permission to use an image commercially. Uncertainty ranges overlap for many entries; their order is not proof of a meaningful difference.
Full results at Arena ↗Arena
Human preferences between generated video clips.
Text-to-video Overall · All licenses
Source date shown by Arena: 2026-11-07. Checked dates record our review, not a new source release.
Compare duration, resolution and audio separately. Early results can move substantially as votes accumulate. Uncertainty ranges overlap for many entries; their order is not proof of a meaningful difference.
Full results at Arena ↗“High,” “max,” tool access and fallback runs can change both results and cost. They are part of the tested configuration.
A small rating gap with overlapping ranges is weak evidence of a clear winner. Preliminary entries may have fewer votes.
An Arena rating of 1,500 cannot be compared with an Intelligence Index of 50. The scales and underlying tasks differ.
Use known answers, count retries and compare time to a usable result. Long context is capacity, not a guarantee of accurate recall.
Compare access and API costs → · A practical benchmark guide →