Run the same prompt three times below. Then turn the dial and run it again.

Finishing a line of code

function isEven(n) { return n

0 — always the top pick1.5 — anything goes
  1. % 2 === 086%
  2. % 2 == 012%
  3. & 1) === 02%
  4. .isEven()0%
  5. / 20%

Hover or tap a bar's note: % 2 === 0 — the standard answer · % 2 == 0 — fine, looser equality · & 1) === 0 — clever bit-trick — but the bracket is wrong · .isEven() — does not exist in JavaScript · / 2 — wrong: returns a number, not true/false

sample 1return n % 2 === 0;
sample 2return n % 2 == 0;
sample 3return (n & 1) === 0;

The top pick still leads (86%), but the runners-up get real chances. Same prompt, different code each run — that's not a bug, it's the dial.

The candidates and samples are a hand-built illustration, not a real model run; the bar maths is the real formula (softmax with temperature). This dial is a picture of what happens inside every model — not a setting you can always change: some APIs no longer expose it (Anthropic's models after Claude Opus 4.6 accept only the default), and even at 0 results aren't fully deterministic.

Different answers from the same prompt. The dial is temperature, and the picking is called sampling.

At each step, the model has a list of possible next tokens, each with a likelihood. It then picks one. That pick is sampling.

Illustrative: possible next words after "const user ="

await

Likelihood (illustrative)
High

getUser

Likelihood (illustrative)
Medium

null

Likelihood (illustrative)
Low

banana

Likelihood (illustrative)
Very low

Anthropic's glossary: higher temperatures lead to more creative and diverse outputs, while "lower temperatures result in more conservative and deterministic outputs that stick to the most probable phrasing."

Even at the lowest setting, don't expect identical output. Anthropic's API reference: "even with temperature of 0.0, the results will not be fully deterministic."

And it's changing. The same reference now marks temperature deprecated: "Models released after Claude Opus 4.6 do not support setting temperature," and only 1.0 is accepted. Checked 25 Sep 2026.

Try it with the rules off. In the sandbox, turn the dial all the way up on the code prompt and watch the variable names get strange. Then try the poem prompt, where variety is the point.

Check yourself

0/3 got it

Saved on this device only. No account, no streaks.

Next in how the AI thinks: Why it invents a library that doesn't exist.