05 · benchmark / games
How good are decision models at Tetris?
autotris hands a hosted decision model one piece at a time, with every legal placement listed. It records the choice, the confidence, the latency, and the cost.
Verified result
20 matches
openai/gpt-6-luna-decisions
17
typesafe-ai/jev
3
openai/gpt-6-luna-decisions beat typesafe-ai/jev 17-3 over 20 matches.
A local sketch: this T, and every hard-dropped resting placement in a fixed well. Seed sketch-t-01, so it comes back the same. The live bench lists every legal placement. Matches are a solo benchmark, model vs model, or human vs model under jstris rules. This sketch is neither a model run nor a jstris match.
34 resting placements · seed sketch-t-01
stem up
stem right
stem down
stem left
Record
seed
sketch-t-01
choice
—
confidence
—
latency
—
cost
—
A hosted model run records confidence, latency, and cost with the choice. This sketch only has a choice.
Solo benchmark
Model vs model
Human vs model
Solo benchmark, model vs model, or human vs model under jstris rules. Every game is seeded and exactly replayable.
runtime
Python >= 3.10
dependencies
None required
models
Vercel AI Gateway
interface
Web UI