home / tools / beach patrol

Beach Patrol.

Every token you send to an LLM burns a turtle. That is not true, but it is what we tell our friends. In this game your prompt is the level: someone is typing it out, every thousand tokens sets a turtle on fire, and you put the fires out before the beach is gone. The turtles are made up. The footnote about what a query really costs is not.

// level settings nothing picked, so you get your last coding-agent session
// exchange rate: source: trust me
// pick a level
0 saved vs 0 lost
0 fires 0.0s best –
🟢 calm 🔥 burning: yellow, then orange, then red 💧 just rescued ⚫ lost meet the turtles ↗
// save a real turtle

Adrián is from Costa Rica, where sea turtles nest on both coasts and a handful of organisations protect them by paying local people to guard beaches instead of raiding them. If the beach above bothered you, these are worth your time or money.

// how the game works, and what tokens actually cost

Your prompt is the level. Someone, somewhere, is typing it out at the top of the board: your own text if you pasted one, otherwise a prompt of truly heroic padding. Every thousand tokens they type (or whatever rate you picked) a turtle catches fire, and the fires come faster as the level goes on. Tap a burning turtle to tip a bucket on it. A fire left alone spreads to the four turtles around it after two seconds and takes the turtle after five, and the turtle you lose floats off with a halo. A fire starts yellow, turns orange, and goes red just before the turtle is gone, so you know which one to run to. One missed fire is a campfire. Two is a wildfire.

Arrow keys and space work too. You win if you save at least as many turtles as you lose. A long coding session is a wildfire you cannot win. That is the point: the only way to have fewer fires is to send fewer tokens.

The turtles are a joke. The exchange rate is made up and you can change it in the level settings. The token count underneath is a real estimate, the same one our LLM Token Estimator uses: roughly 3.5 characters per token for Claude, blended with a word-based cross-check. Expect it to be within 10 to 20 percent of the real number for ordinary prose. If you have an exact figure from a usage page, type it in and it wins.

What a query really costs. Nobody publishes a per-token figure, but two providers have published per-query medians for their chat products. Independent estimates tend to run higher than the providers' own numbers, and long prompts, reasoning modes, and image generation cost more than a short text question.

SourceEnergy per text queryWater per queryRough comparison
Google, technical report on Gemini apps, August 2025 about 0.24 Wh (median) about 0.26 mL under nine seconds of watching TV; five drops of water
OpenAI's CEO, blog post, June 2025 about 0.34 Wh (average) about 0.32 mL an oven running for just over a second; a fifteenth of a teaspoon

Multiply by the hundreds of millions of queries a day and the total is real, which is why data centre siting, cooling, and grid mix are the conversations that matter. Your one prompt is not the problem. Your prompt that repeats the whole document five times is a little bit of the problem.

Full disclosure. This game was built in an AI coding session. Turtles burned to make it: about 300. We regret nothing.

The turtles are sea turtles, or a cartoon of one. The real ones are on Wikipedia, and the organisations in the box above are the ones we would point you to.

Privacy: estimating, the game, and the score card all happen in this browser tab with no dependencies and no network calls. Nothing you paste is uploaded. Your best score stays in your browser. No turtles were harmed.

// a free, browser-only game by Tiempo Software, built in Omaha, Nebraska. Runs entirely in your browser; nothing you enter is uploaded. see all tools.