Token Budget

How Token Budget Works

You are the inference scheduler. Several prompts are queued, each one a needy little task that gets higher quality the more tokens you let it generate. But the context window is finite, the budget is cruel, and every task wants more than its fair share.

1. Each task shows a value curve — quality climbs as you spend tokens, with diminishing returns
2. Use the + / − steppers (or drag the slider) to allocate tokens to each task
3. Your total spend cannot exceed the hard budget meter at the top
4. Hit Submit and the eval compares your total quality against the provably optimal split
5. Clear levels to add more tasks, tighter budgets, and nastier nonlinear curves

The Trick

Different tasks have different value-per-token. The greedy move is to feed whichever prompt currently returns the most quality for its next token — chase marginal value, not raw size. A task with a steep early curve and a hard cap is worth more than a glamorous one that plateaus. Hoarding tokens on a single thirsty agent is reward hacking, and the optimum will notice.