stealthmodels.

StealthMark game benchmarks

Can you plan better than the best AI?

These games test whether an AI model can plan several moves ahead and get better from one level to the next. Every model plays them move by move, with no code and no tools. Play the same levels yourself and get a real benchmark score, ranked among the AI models.

No AI model has played a perfect run yet.

Cooperative planning and learning from experience

Latch

Can you find the best plan and improve it level by level?

Latch shows whether a player can plan a whole sequence before committing to it. Two robots share one goal: one holds a door open on a pressure plate while the other pushes numbered cores onto their sockets. Cores can be pushed but never pulled, so one careless push can strand a core for good.

How to playAmber and Teal are your two robots. Drag one a square, or swipe to move the selected robot. Each move costs a little, each beacon adds a bonus, and Finish ends the level early and pays for every turn left.

5 levels in a row

  1. 1Hold the door
  2. 2Crossfeed
  3. 3Relay vault
  4. 4Passing bay
  5. 5Second thoughts
Watch the AI replays How Latch is judged

How the AI models scored51 models played · median 63.8 across complete runs

  1. 1Claude Opus 5.587.6
  2. 1GPT-6 Astra87.6
  3. 1GPT-6.1 Sol87.6

How your score works

The same levels and the same scoring as the AI.

  1. 1

    The same levels

    You play the exact levels the AI models played, in the same order and with the same number of turns.

  2. 2

    The same calculation

    Each level earns game points. A complete run gets a benchmark score out of 100, worked out exactly like a model’s.

  3. 3

    A place in the ranking

    Finish every level to see where you rank among the AI model runs, then share a link that replays your moves.

More from the benchmarks

Beyond the games

The games are part of StealthMark. The other benchmarks ask models to reason about physics and to draw precise scenes in code.