How long does gpt-5-nano batch take at OpenAI?

85.6x spread — p90 over p50

How much longer the slow tail runs than the typical job. Nearer 1× is more predictable.

Half of gpt-5-nano batch jobs at OpenAI finished within 74s, and nine in ten within 2 hours. That is a spread of 85.6x: the tail sits close to the middle, so what you measure once is close to what you get again. The tightest we measure is gemini-3.7-flash at Google (1.4x); this one sits at 85.6x.

Spread is p90 divided by p50, both measured here over the last 30 days. No provider publishes it; it is the number that tells you whether a median is safe to plan on.

A 85.6x spread is the whole difference between two ways of planning. When the tail sits close to the median, sizing a deadline against the worst case you are likely to see costs you almost nothing over sizing against the typical case — the two numbers are close together. On a model whose tail runs many times its median, the tail IS the plan: the median is a number you cannot safely book against, and the honest planning figure is far out on the right.

That is why the spread matters more than the median alone. This page gives you both, measured, so you can book OpenAI gpt-5-nano against 2 hours and know how often you will beat it.

Every measured job, slowest last

Share of jobs finished, against elapsed time. 3881 completions, 30-day window.

The curve is drawn from the pooled measurements, while the headline percentiles above are the median of each contributor's own figures - so the two need not line up exactly. That is deliberate: no single source can move a headline number by sending more data.

2 hours

Upper end of the measured interval

Plan for 2 hours. That is the upper end of a 90% confidence interval around the p90, computed from the measurements themselves. It is deliberately not the median: being wrong toward synchronous costs twice the money, being wrong toward batch can miss a deadline, and those two mistakes are not equally expensive.

The slowest job we actually observed took 24 hours. That is one job, not a bound.

The spread above is measured over the batch jobs contributors ran in the last 30 days, at the sizes they happened to run. We do not vary job size ourselves, so a much larger or much smaller batch than the ones behind this figure may sit differently — we publish a percentile only for what we have actually timed, and say so plainly rather than extrapolating past it. That is the same discipline that lets you trust the numbers we do print.

Confidence: Low

Score 0.99 · basis: exact_model

Based on a single contributor, a figure pooled across contributors (not yet a per-contributor median).

Pooled across every measurement we have. As more contributors each build their own track record, this figure sharpens into a per-contributor median.

Each measurement is one completed batch job. The duration is end minus start, both timestamped by our server - a client can never submit a duration, and a job we stopped waiting for is recorded as abandoned, never as a fast completion.

Timestamps are absolute and in UTC on purpose: this page may be cached, and a relative age would quietly go stale in the cache while an absolute one stays true.

Same numbers, same delay, no key required.

curl "https://batchwatch.dev/v1/curve?provider=openai&model=gpt-5-nano"

Send five measurements in seven days and the delay is gone: you read every figure live, and the decision endpoints open (/v1/should-i-batch, /v1/estimate-batchtime). Start on the front page.