Latency by band

Median time to first token over the last 24 hours, measured on served requests. Bands set queue admission order, not stream speed: once a response is streaming, every band decodes at the same rate. A higher band buys a sooner start when the pool is busy — so on a quiet pool these numbers converge, and that is the system working, not a mistake.

Latency figures are unavailable right now.

Non-streaming requests have no first-token moment and are excluded rather than counted as zero.

These are observations of a live pool, not commitments — they move with whatever traffic the pool happens to be carrying. What we actually sell is admission: the slots your plan holds open, a bounded wait to be admitted, and a batch deadline we refuse work rather than miss. Those are on the pricing page, and they are the numbers to plan against.