The Best Jev Game Builds So Far

The games category of this directory is where Jev, the decision model from TypeSafe AI, got stress-tested in public first. This is the curated Jev games list: the builds worth your ten minutes, grouped by what each one actually proves. Why games make such an honest test is argued in the real-time games piece; this page is the guided tour, not the thesis.

Every number below is as reported by the builder, linked to its entry. None of it is an official benchmark, because none exists.

Jev games list, part one: the tempo proofs

These are the builds where speed is the whole point, and where a slow model would visibly lose.

• Jev plays Doom. TypeSafe's own launch demo. Game state goes in as structured text (not frames), an action comes back about ten times a second, at roughly $7 an hour as reported by the engineer who built it.
• Slay the Spire 2. Paul Wei had tried a frontier model before and found it slow; with Jev he reports 0.7 seconds per move, faster than he could follow.
• Tetris on the hardest mode. About 0.3 seconds per move, 357 pieces placed and 134 lines cleared in two minutes, as reported.
• Subway Surfers, 50 games at once. Max Blade says the whole run cost under a cent.
• Smash Bros. against itself. All four characters, more than 22 million tokens, and "a couple of cents," per the author.

If you only watch one, watch Doom. If you want the number that surprised us most, it's the Smash run: 22 million tokens and the bill still rounds to lunch money.

Games where Jev is the referee, not the player

The quieter and arguably more useful category. Here the model judges rather than plays, which is much closer to what you'd ship in a real product.

• Wordshift, Marcel Pociot's semantic typing racer, where Jev rules on whether your word fits.
• Battle arena: two fighters, hidden portraits, a verdict. Batman vs Iron Man went to Iron Man at 95%.
• Confidence trivia: you answer with a probability and lose by being overconfident against a model built to output probabilities.
• Jev Word Race: navigating a 1,048-word graph using only visible links, reaching the target in 44.5% of 200 author-measured races, for $0.08 total.

That word-race figure is a useful sanity check. It's not a victory lap, and the builder published it anyway.

The honest losses

A list that only contains wins is an ad. These entries earned their place by showing where fast stops meaning good.

• LLM Chess benchmark. Maxim Saplin's published row: 80 games against a random opponent, 8 wins, 50 losses, zero illegal moves, about $0.0015 a game, as reported. Rule-following, yes. Chess strength, no.
• Jev Chess Lab says it plainly: on its own, Jev still blunders pieces.
• Ms. Pac-Man. David Madurzak measured about 650 ms average response and estimates roughly 70 ms would be needed for normal play, so he paused the game between moves and says so.
• Puyo Puyo. Board-only questions were too weak; staging four questions per move got it past a random player. That's a question-design lesson wearing a game costume, and the judge-questions guide is the long version.

The architecture tricks worth stealing

The builds that will age best are the ones where code owns the world and Jev only owns the branch points.

• Mario Never Dies forks the whole virtual machine into four timelines every time Mario dies. Absurd, and a genuinely clever pattern for search.
• heist-one lets Jev judge guards while deterministic code runs the world.
• Pokemon Red keeps routing and arithmetic in code, asks Jev only at branches, and scores its faint predictions with Brier scores.
• minecraft-agent splits planner from controller: a planner decides what to do, Jev turns it into control decisions.

That last split is the same planner-plus-fast-decider shape the cascade architectures roundup compares outside of games.

Frequently asked questions

What is the best Jev game build to start with?

Jev plays Doom, TypeSafe's launch demo, because it shows the tempo claim most clearly: about ten decisions a second at a reported ~$7 an hour. After that, the chess entries give you the honest counterweight.

Is Jev good at games?

Fast at games, not strong at them. Builders report sub-second moves, while the chess benchmark shows 8 wins in 80 games against a random opponent, as reported. The real-time games piece explains why tempo, not mastery, is the demonstration.

Does Jev see the game screen?

In the Doom demo, no: the entry notes Jev reads a text description of the game state, not the frames. Most builds here follow the same pattern of code turning state into structured text first.

Where can I find more Jev game builds?

The full list lives in the directory's games category, reachable from the front page or search. Built one yourself? Submit it with your numbers.