WorldBench (Cultural Agents): leaderboard
Macro-averaged Complete Task Success across eight culture-language settings under a 30-step cap. Success requires both complete evaluator success and preservation of the non-target workspace.
Metric: Macro-averaged Complete Task Success (%). Source: arxiv.org. Saturation forecast: Around July 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 49.2 | #64 |
| 2 | GPT-5 | 48.8 | #105 |
| 3 | Qwen 3 32B | 48 | #449 |
| 4 | GPT-4o | 38.6 | #357 |
| 5 | Gemini 3.5 Flash | 34.6 | #66 |
| 6 | Llama 3.3 70B Instruct | 24.1 | #541 |
| 7 | Llama 3.1 8B Instruct | 12.5 | #1017 |
| 8 | Qwen 3 4B | 11.6 | #834 |
| 9 | EuroLLM-9B-Instruct | 10.8 | #1047 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=worldbench-cultural-agents · How It Works · Data refreshed daily, snapshot 2026-11-05.