Abstract
Models are asked to build a voxel pagoda. We record the number of output tokens consumed before a pagoda appears. Thinking is enabled. Patience is not.
Results
Voxel Pagoda · k tokens · Lower is better
Methodology
- A pagoda is detected when a vision model, asked "is this a pagoda?", says yes on two of three attempts.
- Runs are stopped at 200k tokens and recorded as "still planning".
Limitations
- Some models produce a superb pagoda after 160k tokens of thinking. We do not penalize quality. We penalize the electricity bill.
Results table
| # | Model | Voxel Pagoda |
|---|---|---|
| 1 | GPT-6 Astra | 18k |
| 2 | Claude Fable 5.1 | 22k |
| 3 | Muse Wonder Full | 30k |
| 4 | Kimi K3 | 30k |
| 5 | Grok 4.7 | 37k |
| 6 | GPT-6 Luna | 38k |
| 7 | GPT-6.1 Sol | 49k |
| 8 | Gemini 3.8 Flash | 57k |
| 9 | Claude Opus 5.5 | 64k |
| 10 | Claude Sonnet 5.5 | 68k |
| 11 | Muse Wonder | 73k |
| 12 | Mistral Large 4 | 73k |
| 13 | Claude Haiku 5.5 | 82k |
| 14 | Qwen3.8 Max | 94k |
| 15 | DeepSeek V4.1 Flash | 102k |
| 16 | GLM-5.3 | 126k |
Citation
@misc{tmb2026voxelpagoda,
title = {Tokens-to-Pagoda: Output tokens consumed before it's a pagoda},
author = {Bro, Trust Me and {et al.}},
year = {2026},
howpublished = {\url{https://trustmebrobenchmarks.com/benchmarks/voxel-pagoda/}},
note = {Results not reproducible. Please do not cite.}
}