Solo benchmark
A model plays a seeded game on its own; we log lines, survival, latency and cost for every piece.
Run locallyLive match on this server Watch it in the viewer →
autotris · a decision-model benchmark
How good are decision models at Tetris?
autotris hands a hosted decision model one Tetris piece at a time, with every legal placement listed, and records what it picks, how sure it was, how long it took and what it cost. First up: openai/gpt-6-luna-decisions and typesafe-ai/jev on the Vercel AI Gateway.
Real match · real time
openai/gpt-6-luna-decisions vs typesafe-ai/jev under jstris rules, every move at its real API latency. Luna won 17-3 over 20 matches.
01 · Tracks
One engine, three ways to play.
Model vs model
Two models, the same pieces, lockstep turns and jstris garbage: a KO, or the most attack at the cap, wins.
Run locallyHuman vs model
You against a model in a live jstris-style 1v1, sending and receiving garbage. Hard drops only for now.
Run locally02 · How it works
No keypresses. One decision per piece.
-
01
List
The engine enumerates every legal hard-drop placement of the current piece: each rotation, each column.
-
02
Decide
The model sees the board and the list, picks one placement and gives a probability for every option. Labels are reshuffled each turn.
-
03
Drop
The piece locks where the model chose. Both sides draw from the same seeded 7-bag, so they always get the same pieces.
-
04
Attack
In a duel, line clears send garbage under jstris rules: attack table, combos, back-to-back and perfect clears, with your own clears cancelling incoming garbage first.
Every result records its seed, ruleset hash, harness version and pinned model id, so any game can be replayed exactly.
03 · First live numbers
Early numbers, clearly marked early.
Luna-decisionsopenai/gpt-6-luna-decisions
- Latency, p50
- ≈321ms
- Cost / 100 pieces
- $0.0136
Jevtypesafe-ai/jev
- Latency, p50
- ≈241ms
- Cost / 100 pieces
- $0.0075
Coming soon
Leaderboard
Solo scores and model-vs-model results will be published here once full benchmark runs are in.
Smoke-test numbers, not benchmark results. Each figure comes from a 3-piece smoke test on 2026-10-12, via the Vercel AI Gateway. They show the models answer, roughly how fast and roughly what it costs; they say nothing yet about how well either model plays.