A new programming benchmark built around Zachtronics’ TIS-100 says large language models can do more than complete constrained coding puzzles: in some cases, they can discover programs that improve on a decade of community optimization work. The benchmark’s authors tested 16 models on 55 puzzles and reported 133 new points on the games’ optimization frontiers, including three outright cycle records.
TIS-100 asks players to program a network of small processors. Each node has two registers and room for only 15 instructions, while neighboring nodes must coordinate their exchanges. A valid solution is scored on execution cycles, nodes used and instruction count. Because improving one measure can worsen another, the community leaderboard tracks a frontier of useful tradeoffs rather than a single overall winner.
The evaluation covered all 47 puzzles in the commercial game and eight community tournament puzzles. Each model received the manual, visible test inputs, register state and game feedback, but not community code or exact leaderboard records. Models had one scored attempt per puzzle and between 10 and 45 minutes, depending on difficulty, to find and refine a working program.
The authors said Opus 5.5 and GPT-6 Astra each solved all 55 puzzles and found new frontier points on 11. Opus produced 72 such points, compared with 32 for Astra, while their reported record-proximity results were close enough that a bootstrap interval did not distinguish them. The benchmark also found that merely solving puzzles did not predict optimization performance: one model solved 53 puzzles but found only one new frontier point, while another solved 52 and found five.
To reduce the risk that a clever program only passed the game’s limited built-in cases, headline results were replayed in a fresh game and checked against 100,000 inputs with an independent simulator. One Opus program for the tournament puzzle Undecimal Twin Rounder completed the task in 1,918 cycles, 29% fewer than the frozen community record of 2,694, although it used one additional node and 28 more instructions. The other two reported cycle records also came from tournament puzzles rather than the main game.
Most gains were new compromises rather than unqualified records. Of the 133 frontier points, 18 beat a community record outright and 115 offered a balance of cycles, nodes and instructions not present in the frozen leaderboard. New points appeared on 14 hard and four medium puzzles, but none of the easy puzzles.
That distinction is important for interpreting the result. The exercise does not establish broad software-engineering ability, but it offers a reproducible environment where generated programs can be executed, stress-tested and compared against years of human work. It also exposes different model strengths in finding any correct solution and in repeatedly refining one under tight resource constraints.
