TALES: leaderboard
TALES (Text Adventure Learning Environment Suite, Microsoft Research): LLM agents play synthetic and human-written text adventures from TextWorld, TextWorld Express, ALFWorld, ScienceWorld and Jericho. The overall score is the average per-game score as a percentage of each game's maximum score (0–100, higher is better); reasoning-effort and thinking-mode runs are listed separately.
Metric: Score (% of max). Source: microsoft.github.io. Saturation forecast: Estimated already saturated. 39 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 (Medium) | 58.7 | #132 (O3) |
| 2 | O3 (High) | 58 | #132 (O3) |
| 3 | O3 (Low) | 54.8 | #132 (O3) |
| 4 | Claude 3.7 Sonnet (Thinking) | 52.5 | #259 (Claude 3.7 Sonnet) |
| 5 | Claude 3.7 Sonnet | 52.1 | #259 |
| 6 | GPT-4.1 | 49.9 | #258 |
| 7 | O1 | 44.2 | #190 |
| 8 | GPT-4o | 40.6 | #357 |
| 9 | Claude 3.5 Haiku | 39.6 | #571 |
| 10 | Llama 3.1 405B Instruct | 36.4 | #470 |
| 11 | Gemini 2.0 Flash | 35 | #351 |
| 12 | Qwen 3 32B | 34.3 | #449 |
| 13 | Llama 3.3 70B Instruct | 32.8 | #541 |
| 14 | Llama 3.1 70B Instruct | 32 | #567 |
| 15 | Qwen 2.5 72B Instruct | 30.7 | #469 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=tales · How It Works · Data refreshed daily, snapshot 2026-11-05.