The answer
The batch endpoints at OpenAI, Anthropic, Google cost 50% of the synchronous rate on both input and output tokens. That is the providers' own published price, not a batchwatch figure - a job that costs $100 synchronously costs $50 in batch.
The reason people hesitate is the wait, so here is the part nobody else publishes: we measured a median of 30s for gpt-5.6-sol at OpenAI. Half price for a wait usually measured in minutes is the whole case for batching.
Every model we have measured
The 50% discount is the same on every model; the wait behind it is not. Most predictable first - spread is p90 over p50.
| Model | Median (p50) | p90 | Spread | n |
|---|---|---|---|---|
| Google gemini-3.7-flash | 2 min | 3 min | 1.3x | 2414 |
| Anthropic claude-haiku-4-5 | 2 min | 5 min | 2.4x | 2606 |
| OpenAI gpt-5.6-luna | 82s | 10 min | 7.6x | 3392 |
| OpenAI gpt-5-nano | 74s | 25 min | 20.3x | 3899 |
| OpenAI gpt-5.6-sol | 30s | 16 min | 32.4x | 1884 |
Each row is measured over the last 30 days, and links to that model's own page with its full distribution and confidence grade. A model we have not yet measured enough of to stand behind a percentile is left out entirely - never filled in with an estimate.
The full numbers, per model
Questions this answers
- How much cheaper is the LLM batch API?
- 50% cheaper - half the synchronous rate at OpenAI, Anthropic, Google, on every input and output token. And the wait is usually far shorter than the 24-hour ceiling: half of gpt-5.6-sol jobs finished within 30s in our measurements.
See it live
Every number here is measured, and it is all public.
The live dashboard shows every model with its distribution, n and confidence grade. /v1/coverage is the same data as JSON, no key required. 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).