SOTA
WEBVOYAGER · 615 TASKS · 99.19% · 610/615 PASSED · BASELINE 98.5% · RANK 01 · JUN 10 2026WEBVOYAGER · 615 TASKS · 99.19% · 610/615 PASSED · BASELINE 98.5% · RANK 01 · JUN 10 2026
WebVoyager Leaderboard·Rank 01·Jun 10 2026

browser-control

browser-control — a shell-native CDP harness — driving Fable 5.

Run

Om Labs’ submissions to the WebVoyager benchmark — 615 live-website tasks across 15 real sites, run with browser-control. Judged by GPT-5.5 (Alumnium vanilla).

99.19%
SOTANew · Jun 10 2026
610/615 passed
Baseline 98.5%Judge GPT-5.5 (Alumnium vanilla)Model claude-fable-5

Run stats

Duration15h 53mavg 1m 33s
Tokens439.3Mavg 714K
Steps3976avg 6.5

About

Methodology

WebVoyager benchmarks agents that browse real, live websites and return natural-language answers. Example: “Search for women’s hiking boots on Amazon filtered to waterproof, 4★+, size 6.” Other notable entrants include OpenAI’s Computer-Using Agent, Google DeepMind’s Project Mariner, and H Company’s Surfer 2, all of which have reported results on this benchmark.

Introduced by He et al. 2024. The browser-control runs (Fable 5, Opus 4.8) share one harness — browser-control, a shell-native CDP tool — and differ only in the driving model. The Jina run used a separate, custom harness that was not open-sourced.

Note — Judged by GPT-5.5 (Alumnium vanilla) on screenshot + response. 615-task scope (643 minus the 24-task Alumnium-removed set and 4 currently-impossible tasks). Self-reported.

Per-site results

Group by

Tasks · 615

Citation

Attribution

Please cite this page as an Om Labs WebVoyager leaderboard submission. For questions, contact Keon Woo Kim <keon@omlabs.xyz>.

@misc{omlabs_webvoyager_2026,
  title = {Om Labs WebVoyager Leaderboard Submission},
  author = {Kim, Keon Woo},
  year = {2026},
  howpublished = {https://webvoyager.omlabs.xyz},
  note = {Contact: keon@omlabs.xyz}
}