TALES - TextWorld: leaderboard
TALES TextWorld subset: one CookingWorld game per difficulty level 1–10 from the NeurIPS 2018 TextWorld competition, in which the agent finds ingredients and processes them to follow a recipe, earning points for sub-goals. Score is the percentage of the maximum game score averaged over games (0–100, higher is better).
Metric: Score (% of max). Source: microsoft.github.io. Saturation forecast: Estimated already saturated. 39 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 (High) | 100 | #132 (O3) |
| 2 | O3 (Medium) | 100 | #132 (O3) |
| 3 | O3 (Low) | 99.1 | #132 (O3) |
| 4 | O1 | 97.8 | #190 |
| 5 | Claude 3.7 Sonnet | 97.3 | #259 |
| 6 | Claude 3.7 Sonnet (Thinking) | 97.3 | #259 (Claude 3.7 Sonnet) |
| 7 | GPT-4.1 | 95.3 | #258 |
| 8 | Claude 3.5 Haiku | 94.9 | #571 |
| 9 | Llama 3.1 405B Instruct | 90.9 | #470 |
| 10 | GPT-4o | 83.6 | #357 |
| 11 | Mistral Large 2 (Jul) | 82.4 | #490 |
| 12 | Gemini 2.0 Flash | 80.8 | #351 |
| 13 | Qwen 3 32B | 79.5 | #449 |
| 14 | Qwen 2.5 72B Instruct | 76.5 | #469 |
| 15 | Llama 3.3 70B Instruct | 69.6 | #541 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=tales-textworld · How It Works · Data refreshed daily, snapshot 2026-11-05.