TowerBench main-1 — first official benchmark run

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.