BabelArena - DeepPlanning: leaderboard
BabelArena (Kuang et al., 2026) adapts English agent benchmarks to 23 high-, medium- and low-resource languages with the BabelFlow translation and verification workflow; this board covers its DeepPlanning family of long-horizon travel and shopping planning tasks, scored by deterministic rules. Pass^1 in percent, the average single-trial task success over three trials pooled across the 23 languages, at medium reasoning effort; higher is better.
Metric: Pass^1 (%; DeepPlanning (long-horizon travel and shopping planning; deterministic rule scoring, Gemini-3.5-Flash converts free-form plans to the required format); Pass^1: average single-trial task success over three trials, pooled over the 23 languages of BabelArena (7 high-, 8 medium- and 8 low-resource) including English; medium reasoning effort, temperature 0, Gemini-3.5-Flash as the simulated user). Source: arxiv.org. Saturation forecast: Around September 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 (Medium) | 36.2 | #56 (Claude Opus 4.8) |
| 2 | GPT-5.6 Terra (Medium) | 22.5 | #60 (GPT-5.6 Terra) |
| 3 | Gemini 3.1 Pro (Preview) (Medium) | 18.6 | #91 (Gemini 3.1 Pro (Preview)) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=babelarena-deepplanning · How It Works · Data refreshed daily, snapshot 2026-11-12.