PlanBench-XL - EGT Precision: leaderboard
PlanBench-XL (2026) tests long-horizon planning of tool-use agents over 327 retail tasks and 1,665 tools. This board is executed ground-truth datatype precision: the percent of unique datatypes produced by the tool calls the agent executes that belong to the ground-truth datatype set of the task, in the default setting at temperature 0. Higher is better.
Metric: Executed ground-truth datatype precision (%): share of the unique datatypes produced by the agent's executed tool calls that belong to the task's ground-truth datatype set, over the 327 PlanBench-XL retail tasks in the default block-free setting (1,665 tools, retrieval-limited tool visibility, temperature 0); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 91.47 | #94 |
| 2 | Gemini 3.5 Flash | 85.29 | #101 |
| 3 | GPT-5.4 | 72.92 | #123 |
| 4 | GPT-5.4 Mini | 71.25 | #389 |
| 5 | DeepSeek V4 Flash | 65.57 | #258 |
| 6 | Qwen 3 32B | 62.36 | #960 |
| 7 | Llama 3.3 70B Instruct | 59.67 | #1357 |
| 8 | Qwen 3 14B | 47.77 | #1233 |
| 9 | Llama 3.1 8B Instruct | 41.33 | #4246 |
| 10 | Qwen 3 8B | 35.31 | #1830 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=planbench-xl-egt-precision · How It Works · Data refreshed daily, snapshot 2026-11-13.