Tetris Bench
Decision models play Tetris through one typed contract: board in, a probability over legal placements out. Same engine, same seeds, swap the brain. IQ rates the choices. Blitz checks who can answer before gravity moves.
Leaderboard
| # | Brain | Score | ±95% | Lines | Latency | $ / 100 decisions |
|---|---|---|---|---|---|---|
| 01 | Greedy heuristic | 274,207 | ±14,326 | 193 | 0.12 ms | — |
| 02≈ | Dellacherie-style | 269,744 | ±10,020 | 195 | 0.12 ms | — |
| 03≈ | Two-ply search | 267,675 | ±2,019 | 198 | 4.0 ms | — |
| 04 | Jev 1.13 | 81,825 | ±25,834 | 96 | 360 ms | $0.014 |
| 05≈ | DeepSeek V4 Flash | 71,705 | ±19,580 | 92 | 1.3 s | $0.0069 |
| 06 | GPT-5.6 Luna | 41,912 | ±10,133 | 69 | 1.2 s | $0.040 |
| 07 | Qwen 3.7 Flash | 19,355 | ±7,174 | 42 | 760 ms | $0.0049 |
| 08 | GPT-4o mini | 2,655 | ±488 | 10 | 1.4 s | $0.024 |
| 09 | Random legal | 427 | ±25 | 0 | 0.17 ms | — |
Score is the mean final score over 30 public seeds with gravity paused (IQ). ±95% is a t-interval across those seeds. ≈ marks a row whose head-to-head interval with the row above includes zero, so the order between them is not established. Latency is the median completed call. Cost is what the provider charged per hundred decisions. Neither enters the rating. Reading the leaderboard →
Blitz
| # | Brain | Score | Lines | Level reached | Answered in time |
|---|---|---|---|---|---|
| 01 | Greedy heuristic | 274,207 | 193 | 20 | 100% |
| 02 | Dellacherie-style | 269,744 | 195 | 20 | 100% |
| 03 | Two-ply search | 267,675 | 198 | 20 | 100% |
| 04 | Random legal | 427 | 0 | 1 | 100% |
| — | Jev 1.13 | not run | — | — | — |
| — | DeepSeek V4 Flash | not run | — | — | — |
| — | GPT-5.6 Luna | not run | — | — | — |
| — | Qwen 3.7 Flash | not run | — | — | — |
| — | GPT-4o mini | not run | — | — | — |
In Blitz the piece keeps falling while the brain thinks. Answered in time is the share of questions answered before gravity moved the piece. Late answers still play when the placement is still reachable. Blitz is reported on its own and does not change the ranking above.
How it works
Contract
Every brain gets the same question: the board, the piece in play, the hold slot, the next five pieces and every legal placement already simulated one step ahead. It answers with a probability over those placements, optionally a hold and a risk estimate. Nothing else is an action.
Modes
IQ pauses gravity while the brain thinks, up to 20 seconds a move, and rates the choices alone. Blitz lets the piece fall in real time. The deadline is the next gravity step, so a slow answer plays a lower piece or none at all.
Rating
The score is the mean final score over 30 public seeds with gravity paused. The ± is a 95% interval across those seeds. Brains are ranked by that mean. Latency and cost are footnotes and never enter the rating.
Results published before 2026-09-20 (tetris-bench@1 and @2) are withdrawn. What changed →