Causal-Plan-Bench: leaderboard
Causal-Plan-Bench (2026) tests physically grounded embodied planning from video: whether a model can judge preconditions, predict action effects, compose multi-step plans and recover from failures, over 1,200 held-out items in 12 tasks. The overall score is the macro-average of six multiple-choice accuracies and six open-ended scores from a GPT-5.6-sol rubric judge, on a 0-100 scale; higher is better.
Metric: Overall score (0-100): macro-average of the 12 task scores, six four-choice accuracy (%) after answer normalization tasks and six open-ended tasks scored by rubric score (0-100): a blind GPT-5.6-sol judge scores each open-ended answer against a human-verified reference from 0 to 1, times 100; 100 held-out expert-reviewed items per task drawn from egocentric and robot video (EPIC-KITCHENS, Ego4D, HoloAssist and others), the same instruction, visual evidence and answer format for every model, temperature 0.2, mean of three regenerated runs; higher is better. Source: arxiv.org. Saturation forecast: Around 2032. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 43.04 | #6 |
| 2 | Gemini 3.1 Pro (Preview) | 38.18 | #94 |
| 3 | Seed 2.1 Pro | 37.22 | #135 |
| 4 | GPT-5.6 Sol | 36.99 | #30 |
| 5 | Kimi K2.5 | 33.66 | #256 |
| 6 | Qwen 3 VL 8B Instruct | 33.23 | #981 |
| 7 | GPT-4o | 32.58 | #741 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=causal-plan-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.