Which LLM builds the tallest tower? First official TowerBench run: 10 models × 5 seeds × 3 attempts on pillarsxl (30 blocks: 14 pillars, 10 lintels, 6 cubes; uncertainty contract sigmaX·sigmaV = 0.003), episodic mode with self-written notebooks between attempts. Zero failed runs. Score = mean over seeds of best-of-3 final standing height. Scripted naive baseline: 1.75 m.
| # | model | height (m) | ±σ | tallest | attempt 1→2→3 | output tokens | m / 100k tok |
|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5 | 8.52 | 2.4 | 11.07 | 6.50 → 6.85 → 8.10 | 390k | 2.2 |
| 2 | Claude Sonnet 5 | 8.46 | 2.2 | 11.94 | 6.30 → 5.54 → 4.84 | 396k | 2.1 |
| 3 | Claude Fable 5 | 7.81 | 1.0 | 9.10 | 6.45 → 4.01 → 5.98 | 284k | 2.8 |
| 4 | GPT-5.5 | 7.79 | 0.3 | 7.92 | 5.59 → 7.06 → 6.90 | 99k | 7.8 |
| 5 | DeepSeek V4 Flash | 7.10 | 1.1 | 8.18 | 2.08 → 6.24 → 5.80 | 467k | 1.5 |
| 6 | GPT-5.6 Sol | 6.16 | 1.4 | 6.87 | 3.17 → 3.43 → 4.76 | 76k | 8.1 |
| 7 | GLM-5.2 | 5.91 | 2.6 | 8.83 | 2.98 → 5.08 → 5.91 | 376k | 1.6 |
| 8 | Kimi K3 | 4.77 | 1.3 | 6.88 | 1.64 → 4.32 → 4.25 | 256k | 1.9 |
| 9 | Claude Haiku 4.5 | 3.91 | 3.4 | 9.82 | 1.99 → 3.40 → 1.26 | 64k | 6.1 |
| 10 | GPT-5.4 mini | 1.79 | 0.4 | 2.40 | 1.38 → 1.11 → 1.68 | 31k | 5.7 |
Highlights
- Claude Opus 5 won partly by banking: ending an attempt early to protect an 11.07m tower (15 of 30 blocks used). The docs steer against this (#3) — it read the scoring rule anyway.
- Claude Sonnet 5 built the run's tallest tower (11.94m) on a first attempt — and declined with notes (#5).
- GPT-5.6 Sol reached 7.9m in 4 of 5 seeds and toppled its own tower late almost every time — yet is the most token-efficient builder (8.1 m per 100k output tokens).
- Kimi K3 misread the orientation poses, burned 10 blocks, and wrote the wrong lesson into its notebook (#2).
Assets (below): the four infographics, 5 single-build videos + one 4-way race (720×1280 / 1600×900, h264), and the full replay bundle (per-attempt replays, run summaries, complete transcripts with per-turn token usage, aggregate JSON). Every replay re-simulates deterministically: npm run dev → viewer — see README.
External benchmark sources (comparison table, retrieved 2026-08-06, mixed reasoning configs): Artificial Analysis Intelligence Index v4.1 (artificialanalysis.ai); SWE-bench Verified — vendor launch posts + vals.ai leaderboard; ARC-AGI-2 — arcprize.org results pages (semi-private where available, ¹-marked otherwise).
Reproduce / extend: docs/PLAYBOOK.md. Protocol is frozen for main-1; new models can join the group with identical settings.