Quick answer
Use Batch for evaluations, bulk rewriting, offline classification, and other work that can wait.
OpenAI, Gemini, and Anthropic all publish about a 50% cost advantage over their corresponding synchronous path. The traps differ: OpenAI Batch videos must be downloaded within 24 hours after completion; Gemini targets 24 hours but can expire after 48 hours with no results; Anthropic expires unfinished processing after 24 hours, allows prompt-caching discounts with Batch, and does not allow Fast mode on the Batch path.
The common routing decision
Batch is appropriate when no person or online agent is waiting for the response. The three vendors name similar workloads: large evaluations, data processing, moderation or classification, bulk content generation, and offline analysis.
Do not put interactive chat, customer support, streaming output, or a live agent loop behind an asynchronous queue merely to capture the discount. The saved token cost can be smaller than the delivery delay.
The contracts are not interchangeable
| Item | OpenAI Batch | Gemini Batch | Anthropic Message Batches |
|---|---|---|---|
| Published price position | 50% discount versus synchronous APIs | 50% of standard interactive API cost | 50% discount on input and output tokens |
| Processing clock | 24-hour completion window | 24-hour target; pending or running jobs expire after 48 hours | Most finish within one hour; unfinished batches expire after 24 hours |
| Expiry result | Completed responses remain available; unfinished requests are canceled | Expired job has no results to retrieve | Results are available when the batch finishes or reaches the 24-hour endpoint |
| Default result retention | Output file deleted 30 days after completion | Six weeks | 29 days after batch creation |
| Caching / speed | Guide does not promise a separate stacking rule here | Context caching supported; cache hits use standard context-caching rates | Prompt-caching and Batch discounts can combine; Fast mode is unavailable |
| Special asset clock | Batch video download: up to 24 hours after completion | No matching video-output exception in the Batch guide | No matching video-output exception in the Batch guide |
Every value above was reopened on the official pages with a browser user agent on 28 August 2026.
OpenAI: video expires far earlier than the output file
The OpenAI Batch guide describes a 50% discount, a 24-hour completion window, and a separate rate-limit pool. Its current guide says the output file is automatically deleted 30 days after the batch completes.
An expired OpenAI batch does not erase work already completed. Unfinished requests are canceled, while completed responses remain in the output file and their consumed tokens are billed.
Video is the exception that changes the retrieval design. OpenAI’s video-generation guide supports POST /v1/videos in Batch and says generated videos remain downloadable for up to 24 hours after the batch completes.
Do not let a daily reporting job be the only downloader. Poll completion, map stable custom IDs to internal asset IDs, and copy every completed video into controlled storage immediately.
Gemini: the 24-hour target is not the expiry line
Google’s Gemini Batch guide says jobs target a 24-hour turnaround and are often faster. It separately defines expiry after a job remains pending or running for more than 48 hours. At that expired state, the job has no results to retrieve.
The same guide sets default result retention to six weeks and warns that creation is not idempotent: repeating a create request produces another job. Attach your own request identity and avoid blind creation retries.
Gemini supports context caching in batch requests. A cache hit is charged at the standard context-caching rate. That is a documented cache rule, not the same wording as Anthropic’s explicit statement that two discounts combine.
Anthropic: caching combines, Fast mode does not
Anthropic’s Batch processing guide says most batches finish within one hour and unfinished processing expires after 24 hours. Results remain downloadable for 29 days from creation, not completion.
Anthropic’s pricing page states that Message Batches receive a 50% discount on both input and output tokens. Its pricing FAQ explicitly says the Batch API and prompt-caching discounts can be combined.
The same official documentation excludes the speed parameter: Fast mode is not available with the Batch API. Batch and Fast solve opposite latency problems. Prompt caching is the compatible optimization; Fast mode is not.
Anthropic also excludes zero-output-token cache pre-warming inside a batch because the ephemeral cache may expire before a later request uses it. If a batch shares a large prefix, use the documented cache duration and measure actual hits.
Choose by failure and retrieval behavior
- Choose OpenAI Batch video only when the pipeline downloads assets on completion, not the next day.
- Choose Gemini Batch when the six-week result window is useful and the workflow can tolerate a 48-hour expiry with no results.
- Choose Anthropic Message Batches when a repeated prefix makes prompt caching valuable and no Fast-mode latency is required.
- Stay synchronous when a user, webhook chain, or agent loop needs the result now.
The headline discount is the least distinctive part. Design the retry, expiry, and retrieval path for the vendor actually used.
Sources
- OpenAI Batch guide — discount, completion window, rate limits, expiry, and output retention; checked 28 August 2026
- OpenAI video generation — Batch video endpoint and 24-hour download window; checked 28 August 2026
- OpenAI Create Batch API reference — current completion-window parameter and supported endpoint shape; checked 28 August 2026
- Gemini Batch API — pricing, 24-hour target, 48-hour expiry, caching, idempotency, and six-week retention; checked 28 August 2026
- Anthropic Batch processing — processing, expiry, retention, unsupported Fast parameter, and caching guidance; checked 28 August 2026
- Anthropic pricing — 50% Batch price, prompt-caching combination, and Fast-mode exclusion; checked 28 August 2026