Sixteen LLMs find new TIS-100 score frontiers on 18 of 55 puzzles
Original title: LLMs Play Zachtronics' TIS-100
Opus 5.5 and GPT-6 Astra each solved all 55; Opus found 72 new frontier points to Astra's 32.
Researchers gave 16 large language models Zachtronics' TIS-100, its manual and coarse score histograms, but not community code or exact records. Each model had one timed attempt on 47 game puzzles plus eight tournament puzzles. The run produced 133 new frontier points and three outright cycle records; players later improved most of the programs. A counted point required a fresh-game replay and 100,000 passing tests.
Unconfirmed: The proximity gap's 95% bootstrap interval includes zero, so the data does not separate Opus 5.5 and GPT-6 Astra.
- Time limits were 10, 30 or 45 minutes depending on puzzle difficulty.
- Record proximity was 78.0 for GPT-6 Astra and 77.2 for Opus 5.5.
- On Undecimal Twin Rounder, Opus 5.5 finished at 1,918 cycles versus a 2,694-cycle record.
- Gemini 3.8 Flash solved 53 puzzles but found new points on only one.
- MiMo-V2.6-Pro solved 10 of 55 puzzles and found no new frontier points.
- Products
- TIS-100
- Organizations
- Zachtronics
From the original article
16 min read at tis-100.wincon.dev
On this page Leaderboard How it works One record up close What separates the models What the models built God mode Co-discovery Pareto explorer Methods Related work Cite Credits All 55 puzzles → Twelve nodes. Two registers per node. Fifteen lines of code.
By fairestvalue · tis-100.wincon.dev
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.