Half the price, measured wait
Half the price. Here is the real wait at Anthropic.
Anthropic's batch API costs 50% of its synchronous rate - the provider's own published number, on every input and output token. The only thing it does not tell you is the wait, and Anthropic publishes just one figure for that: a 24 hours completion window, the ceiling in its contract, not a measurement of the queue.
We measured it. Half of claude-haiku-4-5 jobs finish within 3 min and nine in ten within 7 min, across 2588 completed jobs - far inside the 24 hours Anthropic quotes. That is the number worth planning against; the published window is the fallback until each model has the sample to beat it.
Measured models
30-day window. A dash marks a model still building its baseline: we publish a percentile from 2 measurements up, and holding that line is what makes every figure here trustworthy.
| Model | p50 | p90 | p95 | Jobs | Sources | Confidence | Last measured |
|---|---|---|---|---|---|---|---|
| claude-haiku-4-5 | 3 min | 7 min | 10 min | 2588 | 0 | Confidence: Low | 2026-10-09 |
Questions this page answers
- How long does the Anthropic batch API take?
- For claude-haiku-4-5, half of jobs finished within 3 min and nine in ten within 7 min, across 2588 completed jobs - a measured answer no provider publishes. The rest of the table is building its own baseline right now, and each model gets the same treatment the moment its sample is ready.
- What does Anthropic itself promise?
- A completion window of 24 hours. That is a contract they wrote, not a measurement - a ceiling everyone quotes because nobody has timed the real thing. We have, and the measured figures above are the ones worth planning against; Anthropic's window is the dependable fallback until a model has the sample to beat it.
- How current are these numbers?
- The newest measurement behind this page finished at 2026-10-09T22:07:39.000Z. Public figures are precomputed and delayed by 15 minutes; anyone contributing measurements reads them live, with no delay at all. Every row states its own age rather than relying on this page being fresh.
This question has its own measured answer page: How long does the Anthropic Message Batches API take?
Get this as JSON
No key required for coverage.
curl "https://batchwatch.dev/v1/coverage"
Per-model curves live at
/v1/curve?provider=anthropic&model=....
The front page has the decision endpoints.