Complexity Ceiling Benchmark: leaderboard
Sequential state tracking under growing depth (Complexity Ceiling Benchmark, Chapra et al., 2026): procedurally generated tasks in a spatial grid, a symbolic pointer register and a transitive social-relations graph, each at depths 5 to 50 with 40 seeded trials per depth. A trial counts only if every intermediate step and the final state are exact; accuracy in percent averaged over the three domains; higher is better.
Metric: Accuracy (%; share of parseable trials whose final answer and every intermediate step are exact, 40 seeded trials at each depth N = 5, 10, ..., 50; temperature 0 via OpenRouter; mean of the three domain accuracies, 1,200 trials per model). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 31.9 | #455 |
| 2 | Gemini 2.0 Flash (001) | 18.4 | #724 |
| 3 | Llama 3.3 70B Instruct | 1.9 | #1353 |
| 4 | GPT-4o Mini | 1 | #1627 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=complexity-ceiling-benchmark · How It Works · Data refreshed daily, snapshot 2026-11-12.