EN中
← All visual benchmarksArena / 2026-10-20
Arena · AGENTS

Agent capability · IPS

Inverse propensity score; not a percentage.

Hover or tap a bar to inspect. Select up to 3 models to highlight and compare.
Model scores10 models
Agent capability · IPS. Zoomed horizontal axis, from 0.0444 to 0.2307ZOOMED X-AXIS · 0.0444 – 0.2307 · NOT FROM ZERO0.04440.09100.13760.18410.2307Claude Opus 5.5 (Max)Claude Opus 5.5 (Max) · 0.2092 IPS · zoomed horizontal axis0.2092GPT-6 Astra (Max)GPT-6 Astra (Max) · 0.147 IPS · zoomed horizontal axis0.147Claude Fable 5.1 (Max)Claude Fable 5.1 (Max) · 0.1417 IPS · zoomed horizontal axis0.1417GPT-6.1 Sol (Max)GPT-6.1 Sol (Max) · 0.1355 IPS · zoomed horizontal axis0.1355Claude Sonnet 5.5 (Max)Claude Sonnet 5.5 (Max) · 0.1267 IPS · zoomed horizontal axis0.1267Gemini 4 Argon (High)Gemini 4 Argon (High) · 0.1203 IPS · zoomed horizontal axis0.1203GPT-6 Sol (Max)GPT-6 Sol (Max) · 0.0936 IPS · zoomed horizontal axis0.0936Claude Fable 5 (High)Claude Fable 5 (High) · 0.0831 IPS · zoomed horizontal axis0.0831Claude Opus 5 (Max)Claude Opus 5 (Max) · 0.0809 IPS · zoomed horizontal axis0.0809GPT-5.6 Sol (xHigh)GPT-5.6 Sol (xHigh) · 0.0659 IPS · zoomed horizontal axis0.0659
Choose a bar for details.

How to read this chart

Official top-ranked excerpt. Human preference ratings are not accuracy percentages; IPS is a distinct metric.

Score unit: IPS. Values can only be compared within the same task and protocol.

Source publication: 2026-10-20. Higher values indicate better results for the reported metric.

Source: Arena · CC BY 4.0. Redrawn from released data; reduced to selected models and rounded. Not affiliated with the source organization.

← Explore all benchmark charts