AI models beat records in TIS-100 programming game

Zachtronics' TIS-100 asks players to write code across a grid of small nodes. Each node behaves like a tiny processor, and neighboring nodes must wait for one another to exchange values. Even multiplying two numbers becomes a small engineering project, and the community has spent a decade optimizing these programs on a public leaderboard.

According to the report, a team gave 16 large language models the game and its manual, but withheld the community leaderboard and exact records. The models were asked to find a working program and keep improving it. Each model attempted all 47 main puzzles plus eight tournament puzzles in a single scored attempt, with 10 to 45 minutes per puzzle depending on difficulty.

The article says the models picked up the game within days, and community players soon took the models' programs and improved most of them. Opus 5.5 and GPT-6 Astra led the field, solving all 55 puzzles, and both found new frontier points on 11 of them. But solving and optimizing turned out to be different skills: Gemini 3.8 Flash solved 53 puzzles but found only one new point, while Fable 5.1 solved fewer, 52, yet found five new points.

One tournament puzzle produced the most striking result. According to the report, Opus 5.5 cut the old community record of 2,694 cycles down to 1,918, a 29% faster solution, though it cost one more node and 28 more instructions. The team says every result is a program that can be run and tested on fresh inputs, and it required a fresh-game replay plus 100,000 passing tests for its headline count.