Batch inference trades latency for price. You submit a file of requests, the provider processes them without an interactive deadline, and the results arrive later, often within hours. Several major providers price this at half the standard rate on input and output, which makes it the cheapest lever for work that does not need an immediate answer.
It is the right tool for offline evaluation, backfills, classifying a stored corpus, and generating reports. It is the wrong tool for a chat, a support widget, or an agent loop with a person waiting. Comparing a batch price with an interactive price is comparing two products.
Batch jobs can often combine with prompt caching when the requests share a prefix, and the same cost arithmetic applies. Record the turnaround you actually observe, not the SLA, and keep the job idempotent so a retry does not double-write. See prompt caching.