SpaceChem Bench: Can LLMs Beat Humans at 2D Spatial Programming?

We ran 16 LLMs, covering leading frontier and open-weight models. Every model attempted every non-boss puzzle in the main campaign, the 63 Corvi expansion, and the three Team Fortress 2 levels, in one scored attempt each (pass@1), with 10 to 60 minutes per puzzle, depending on difficulty. 37 of the puzzles run inside a single fixed reactor. In the other 18 (SpaceChem’s production levels), the model routes pipes between tanks and reactors on a terrain map, from a zero-reactor pipeline tutorial up to six-reactor chemical plants.

SpaceChem has no text format to type into. Players drag arrows and instructions onto the grid with a mouse. We run the unmodified Linux release of the game on a virtual display and give models text tools instead. A model’s edits are written straight into the game’s save file, and every score in this article comes from the real game.

A passing solution is only the starting point: models keep editing and running until the 10- to 60-minute clock expires, trading symbols for cycles or recovering an earlier checkpoint with get_best when a rewrite crashes. On No Employment Record Foundresearch_6, for example, Sonnet 5.5’s first passing reactor two minutes in took 1,113 cycles; twenty minutes later it reached 465 cycles before finishing at 464/1/23 and 504/1/22 (examined up close below).

In TIS-100 Bench, Opus 5.5 and GPT-6 Astra finished in a statistical tie. Moving from 1D assembly to a 10×8 spatial grid separates them: GPT-6 Astra leads at 74.3 record proximity for an estimated $5.55 per puzzle, while GPT-6.1 Sol takes second at 69.7 for $1.17, 4.7 times cheaper than GPT-6 Astra and 4.2 times cheaper than Opus 5.5 (66.2 at $4.90). Muse Wonder 1.0 is the most expensive model tested at $9.94 per puzzle and finishes ninth at 36.8.

Models ran their reactors 23,745 times, and 32% passed. Of the 16,248 failed runs, 57% crashed from physical collisions: two atoms meeting in one cell, a rotated molecule clipping a wall, or a new input spawning onto an atom that hadn’t cleared the intake zone. Another 26% deadlocked in loops that never reached the target count, and only 12% actually output the wrong molecule (5% failed static checks before simulation). Models usually figure out the chemistry; they lose on 2D traffic control.

Across all 16 models, 28 model-frontier points beat the frozen archive; pooling across models deduplicates shared discoveries and drops six points superseded by another model on No Employment Record Foundresearch_6, An Introduction to Sensingsensing_1, and Ice to Meet Youfusion_1, leaving 20 nondominated points across 11 puzzles. None of the 20 beats the community’s best cycle, reactor, or symbol count outright (each is a new tradeoff), and 11 of them knock an existing community solution off the frontier by matching or beating it on all three scores. All 20 are on single-reactor puzzles.

SpaceChem Bench started with a challenge on the Zachtronics Discord. After early LLM benchmarks came out on Opus Magnum and TIS-100, @12345ieee, who holds hundreds of records in the SpaceChem archive, asked when models would come to SpaceChem:

When our runs finished, two of the models’ new frontier points beat @12345ieee’s own solutions in the frozen archive: 465/1/24 on No Employment Record Foundresearch_6 and 314/1/10 on Moustachium 604tf2_2. We submitted selected frontier solutions to the community archive on October 27, including Sonnet 5.5’s, GPT-6 Astra’s, and Opus 5.5’s four new points on No Employment Record Foundresearch_6.

Three days later, @12345ieee came back and posted 463/1/23, 462/1/24, 503/1/22, and 544/1/21 on No Employment Record Foundresearch_6, each one to four cycles faster than the model solution it replaced. The new files share most of their lines with the model files: 463/1/23 keeps Sonnet 5.5’s layout, moving one red bond instruction and shifting a three-arrow corner by one cell (4 of the 29 MEMBER lines).

Just like on TIS-100, neither side got there alone. The models found new layouts and symbol tradeoffs across the archive, and players read those reactors, saw where a corner or instruction could tighten by a cycle, and pushed the frontier further.

Grouped by main mechanic. Select a puzzle to open it in the explorer, or browse all 55 puzzle pages.

TIS-100 Bench is our companion benchmark on Zachtronics’ assembly game, using the same 16 models, 55-puzzle pass@1 structure, and Pareto-archive evaluation. Where models solved 82% of TIS-100 episodes, found 133 new frontier points, and broke three cycle records outright, the shift to 2D waldo routing and multi-reactor pipelines in SpaceChem cut overall solve rate to 56%, yielded 20 new tradeoff points (with no outright 1D records), and separated models that tied on 1D assembly. Opus Magnum Bench evaluates agents on a third Zachtronics puzzle game using a single-solution sum score (cost + cycles + area), which we include alongside record proximity, 3D hypervolume, and frontier coverage.

Thanks to Zachtronics, the SpaceChem players, and the maintainers of the community archive, SChem, and the leaderboard bot. Their solutions and tools make this comparison possible. Thanks also to 12345ieee for taking the models’ solutions further.

SpaceChem, TIS-100, and Opus Magnum are trademarks of Zachtronics. Team Fortress 2 is a trademark of Valve. Model names and logos are trademarks of their respective owners and appear only to identify the models tested. Wincon Research is not affiliated with or endorsed by Zachtronics, Valve, or any model provider.