Tetris as a legal-move test: Jev picks from the landings you give it
Two independent demonstrations evaluated Jev playing Tetris from enumerated legal placements, comparing latency and costs against Claude Haiku 4.5 and Gemini 3.5 Flash-Lite.
On September 17, Ashutosh Mathore shared a comparison of three models playing the identical 200-piece Tetris sequence, selecting moves from an identical list of valid landings:
- Jev: 9,200 points, 75 lines, ~300 ms per move, 0 illegal moves
- Claude Haiku 4.5: 8,900 points, 72 lines, 1.52 s per move, 0 errors, $0.48
- Gemini 3.5 Flash-Lite: 9,000 points, 75 lines, 1.13 s per move, 0 errors, $0.02
Mathore noted that Jev’s competitive score reflected lower per-decision latency rather than superior game strategy.
The following morning, Steve shared a separate demonstration where Jev selected among valid placements at roughly 400 ms per move, costing approximately one cent per 100 pieces, noting that schema constraints prevent illegal moves. Reviewing a similar run, Addison observed that while Jev does not demonstrate advanced gameplay heuristics, it executes valid moves consistently.
Like the browser demos, the game loop limits decisions to valid moves: the program calculates legal piece placements, Jev selects an option, and the engine executes the move. Tetris offers a straightforward test of this pattern because all legal placements can be enumerated and verified before each choice.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Ashutosh Mathore posted a scoreboard comparing models across a single 200-piece run: Jev scored 9,200 points, cleared 75 lines, averaged ~300 ms per move, and made 0 illegal moves; Claude Haiku 4.5 recorded 8,900 points, 72 lines, 1.52 seconds per move, and cost $0.48; Gemini 3.5 Flash-Lite recorded 9,000 points, 75 lines, 1.13 seconds per move, and cost $0.02. Steve reported roughly 400 ms per move and one cent per 100 pieces at normal playback speed. We did not re-run the benchmark code. The absence of illegal placements follows directly from constraining model choices to valid engine states.
Compare
This structure parallels the Browser Use and Stagehand browser workflows: the host application calculates legal candidate actions, Jev selects an option, and the runtime executes the move. Tetris isolates this pattern within a strictly bounded problem space. Minor score differences over a single piece sequence remain inconclusive, but the latency contrast (300 ms compared to 1.1 to 1.5 seconds) matches the speed characteristics emphasized in TypeSafe documentation.
Terms
- Legal landing
- A valid board placement calculated by the game engine. The model chooses exclusively among these options.
- Action space
- The set of valid moves available for the active piece, recomputed on each turn similarly to how browser agents rebuild interactive DOM elements.
- Schema
- The structured Choice definition representing possible piece landings, preventing the model from outputting invalid moves.