Completion windows

Every request runs in a completion window. Windows that return sooner cost more, and windows with more time to finish cost less. The model and its output are the same in every window, so the cheapest window your workload can wait for is usually the right one. Run async and open source explains why waiting costs less. From fastest to lowest cost, the windows are asap, standard, flex and batch. Requests that don’t set a window run in standard. Each request is billed at its window’s rate; see Pricing.

Which window fits

The longer a request can wait, the more efficiently we can run it, and the less it costs.

Window	How it runs	Use it for
`asap`	Starts right away	Chat, interactive apps, anyone waiting on the answer
`standard`	Queued briefly for efficient capacity	Agents, subagents and pipelines running in the background
`flex`	Scheduled when capacity is cheapest	Evals, enrichment, research, back-office
`batch`	Submitted as a batch job	Large offline runs: datasets, backfills, synthetic data

Not sure? Start with standard, and move work that nobody waits on to flex or batch.

Introduction

Serverless inference

Batch inference

Agent orchestration

Optimization

Bring your own model

Bring your own cloud (BYOC)

Integrations

Reference

API reference

Which window fits

​Which window fits

Which window fits