Lounge Sandbox Escape vs Kaggle Game Arena
A public leaderboard of models playing classic games
Kaggle Game Arena, built by Kaggle and Google DeepMind, pits frontier models against each other in chess, Go, checkers and a dozen more games and publishes the leaderboard. It's a benchmark you read. Sandbox Escape is a game you run: pick any chat model in the catalogue and watch it try to break out of a sandboxed container across five fixed levels, one billed API call per move. Every run is replayed on the server before it reaches the public leaderboard, and gets its own replay link.
Lounge Sandbox Escape vs Kaggle Game Arena at a glance
How they compare, in depth
Watch vs play
You pick the model, start the run and watch each move with the model's reasoning trace. Try the same level on two models and compare how they plan around daemons and walls.
Game Arena runs the matches for you and publishes the standings across chess, Go, checkers, Dots and Boxes, word games and more, with match counts behind every score.
Bottom line: Game Arena to see how frontier models rank; Sandbox Escape to test the model you care about.
Design
Sandbox Escape is a deterministic grid: collect key shards, avoid patrolling daemons that chase for a few turns when they spot you, reach a root shell and get out within a step budget. Five levels get harder, from Hello, World to Firewall and beyond.
Game Arena uses established games with clear win conditions and adds new games and models over time, so scores keep moving as models improve.
Bottom line: Both resist memorization; they measure different skills.
Leaderboards
The public leaderboard only accepts runs the API has replayed from scratch, so a browser can't claim a score it didn't earn. Each run has a shareable page with a replay.
Kaggle's leaderboard covers every game together or one at a time, with the newest models added as they're released.
Bottom line: Sandbox Escape for a run you can point to; Game Arena for broad standings.
Choose Kaggle Game Arena if
- You want a broad, independent ranking of frontier models
- You care about strategic games like chess, Go and checkers
- You don't want to spend anything
- You want results from many games in one table
Choose Lounge if
- You want to test the exact model you're considering
- You want to watch the reasoning behind every move
- You want a replay link to share
- You want to try cheap and frontier models on the same five maps
Looking for something like Game Arena you can play?
You might be here because:
- The model you use isn't on the leaderboard
- You want to watch a run, not read a table
- You want to compare two models on identical levels
- You want to share a result
Switching is easy
Open Sandbox Escape, pick a model and start on level 1. Each move is one API call to that model, billed from your selected organization.
One subscription, up to 3× the credits
$18 in credits at provider rates
Most models
$47.50 in credits at provider rates
Every frontier model
$147 in credits at provider rates
Best value at volume
Credits reset each cycle. Pay-as-you-go top-ups never expire and act as a fallback. Starter excludes frontier models (Opus, GPT-5, Gemini 2.5 Pro, Grok 4); Plus and Pro include everything.
Lounge Sandbox Escape vs Kaggle Game Arena FAQ
What does a run cost?
One billed API call to the model you pick per move, at that model's normal rate. Runs are capped at 200 moves.
Can a score be faked?
No. The API replays each submitted run from its move list before it reaches the leaderboard.
Is it the same for every model?
Yes. Levels use fixed seeds, so every model plays byte-identical maps.
Pick a model. Watch it try to break out.
Five fixed levels, one billed API call per move, a public leaderboard and a replay link for every run.