Breaking01 / AI
WebArena-Pro evaluates audio and video web agents

The researchers behind WebArena-Pro reported a joint top score of 62.3% for Codex and Claude Code (n=300, WebArena-Pro, different model pairings) in a paper posted to arXiv on 16 Nov 2026, evaluating audio and video tasks in isolation; the preprint is not peer-reviewed, and its abstract reports no independent replication.
The researchers report that their self-hosted web applications treat audio and video as observations alongside text and images. Dedicated actions let agents extract information from audio and video content.
According to the abstract, the shared highest score comes from configurations that pair each agent with a different model. The result therefore describes those agent-and-model combinations under the benchmark’s evaluation conditions.
The researchers report that tasks are created through a dedicated annotation interface, screened with language-model assistance and then validated by people. Each task runs in isolation, and the benchmark supports parallel evaluation.
The authors argue that earlier reproducible benchmarks cover only a small range of web applications and limit observations to text and images. Evaluations using live websites offer broader coverage, but changing pages and data make results harder to reproduce, they argue.
The abstract does not report task-level scores or score variation across repeated runs. Those omissions leave the distribution and consistency of the reported performance unspecified.