What you're looking at
This is not an illustration of the response. As the model writes, it computes a full
probability distribution over its entire vocabulary (~100,000+ possible next tokens) at every single step —
how confident it is, and everything it almost said instead. That distribution is normally thrown away the
instant a token is chosen. This page intercepts it, via the API's logprobs feature, and turns
it directly into pixels and sound before it disappears. No human judgment, no camera, no traditional
instrument can perceive next-token entropy — without the model actively running, there is nothing here
to render.
The honest limits: the API only reports the top 5 candidate tokens per step, not the full ~100k-wide distribution. The entropy shown, H = −Σ p·log₂(p), is computed only over those 5 — so it is a systematic under-estimate of the true entropy (the hidden remaining probability mass, however small, still carries positive entropy that this number cannot see). See the self-test suite above — one test constructs this exact case and asserts the direction of the bias.
DEMO replays a trace captured from one real streaming API call, at the real recorded token timing (network jitter and all) — nothing here is faked, but it is not happening right now. LIVE is the model generating in front of you this instant, using your own key. The two are visually flagged at all times so it's never ambiguous which one you're looking at.