Grid-Bench T1 pooled exact: leaderboard
Grid-Bench multidimensional cellular-automaton state tracking, prompt 0.4; September16–24 collection, one response with 128000 combined tokens, no tools, 7200-second deadline. High effort for five models; MiniMax M2.5 native always-on thinking. Exact-output counts retain invalid responses and all failures. NAIRR routes for GPT/Claude, native Bedrock for GLM/Kimi/MiniMax; Appendix C pins identities. 1 updates pooled over dimensions 2/3/4, 192 main instances; first count of exact/incorrect/invalid tuple. Metric: Exact answers (out of 192). Higher is better. Source: https://arxiv.org/html/2610.22312v1
Metric: Exact answers (out of 192). Source: arxiv.org. Saturation forecast: Around January 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (High) | 183 | #18 (GPT-5.6 Sol) |
| 2 | MiniMax M2.5 (Thinking) | 150 | |
| 3 | Kimi K2.5 (High) | 148 | #144 (Kimi K2.5) |
| 4 | Claude Sonnet 5 (High) | 137 | #54 (Claude Sonnet 5) |
| 5 | GPT-5.4 Mini (High) | 69 | #215 (GPT-5.4 Mini) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #7 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=grid-bench-t1-pooled-exact · How It Works · Data refreshed daily, snapshot 2026-10-17.