Live Claude Batch API turnaround times
Checking the latest batches…
- Latest batch took
- —
- Oldest still waiting
- —
- Last batch sent
- —
Timetable for —
- Typical wait, 24 h
- —
- 95% back within
- —
- Slowest, 24 h
- —
- Batches timed
- —
Every batch in the last 24 hours
came backstill waitingfaileddelayed, 10 min or more
The last 7 days, hour by hour
typical waitup to the 95th percentile
Recent departures
| Sent | Took | Status |
|---|
How it's measured
Every 15 minutes batchtimes sends a real batch to the Claude Message Batches API: one request asking the model to reply “OK”. Each batch is timed from the API's own created_at to its ended_at, so the figure is what the API reports, not how often we happen to look.
A batch only ends when its slowest request does, so every model gets a batch of its own. A model shows as not running when we haven't sent it a batch in the last hour.
batchtimes is run by Navanto, which uses the Batch API every night to write AI risk briefings for construction projects. We built it to find out whether a slow batch was us or everyone.
Questions
How long do Claude batches usually take?
Anthropic says most batches finish within an hour, and every batch finishes or expires within 24 hours. The board above shows what a small batch is getting right now.
Why is my Claude batch taking so long?
Batches seem to finish in waves: several end within the same few seconds, then nothing for a while. Most of the wait is spent queued for a wave, not processing, and now and then one batch is passed over by several waves while identical batches sent after it come back. If the board shows delays too, the queue is slow for everyone.
Does a bigger batch take longer?
Not in our tests so far. A batch with 22,000 input tokens and a long answer came back in the same wave as a one-word request, and so did a batch of ten requests. We're still timing them through busy periods.
Is batchtimes run by Anthropic?
No. It's an independent site that times the public API like any other customer.