The Batch API processes up to 50 thousand asynchronous requests with a 24-hour window. See where the model helps and what business controls are needed.
Direct answer
OpenAI's Batch API receives JSONL files with large volumes of requests and processes them asynchronously, with a current window of up to 24 hours and a reported discount of 50%. The documentation supports up to 50 thousand requests per file, limited to 200 MB, and associates each result with the custom_id defined by the client. For businesses, it is suitable for assessments, classification, enrichment and offline generation; does not replace interactive flows or eliminate reconciliation, privacy, and failure handling.
Asynchronous is a product decision
The queue works when the result can arrive later: catalog classification, periodic evaluations, embeddings and base enrichment. Real-time service continues to require a synchronous route.
custom_id is the reconciliation key
The order of answers should not be assumed. A unique, traceable identifier links each output to the source record and allows retry only failures.
Economy depends on useful occupation
The discount does not compensate for duplicate entries, excessive prompts, or discarded outputs. Deduplication, sampling, and size limits must precede submission.
Partial failure needs policy
A batch may conclude with unsuccessful items. The operation must separate transient errors, invalid content, and permanent failures before reprocessing.
Nexus Reading
Batch is an operational layer for AI at scale, not an architectural shortcut. Value comes from combining queuing, prompt versioning, auditing, and objective quality criteria.
FAQ
How long can a batch take?
The currently documented window is up to 24 hours; therefore, it is not suitable for interactive responses.
How many requests can fit in one file?
The reference reports up to 50 thousand requests and a maximum size of 200 MB per input file.
Do the results arrive in the same order?
The application must use the custom_id to reconcile each response, without depending on the order of the output file.
Essential guides to delve deeper into the decision
This editorial analysis was produced by Nexus from the official sources below, consulted on September 20, 2026. The text is original and interprets practical implications for companies.
- OpenAI — Batch API reference: limits, processing window, states, API input and output formats.
- OpenAI Cookbook — Batch processing: official example of JSONL preparation, sending, monitoring and association of results.
Date reported by main source: official documentation consulted on September 20, 2026; technical analysis.