Last updated: October 2026
- All three batch APIs charge 50% of the standard price for work that can wait up to 24 hours.
- Request caps differ sharply: Claude takes 100,000 requests or 256 MB, OpenAI 50,000 requests or 200 MB, Gemini 2 GB per input file.
- Results come back in any order on Claude and OpenAI, so every request needs a custom_id you can join on.
- Retention varies: Claude keeps results 29 days, OpenAI output files 30 days, Gemini 6 weeks.
A bulk SEO pipeline does not need a title tag in two seconds. It needs 3,000 of them for the lowest bill.
Meta descriptions, FAQ blocks, product copy and internal link suggestions all tolerate a wait. That is the workload every provider discounts by half.
How Do the Three Batch APIs Compare on Limits and Price?
The discount is identical. The ceilings are not, and they decide how many batches you have to split a site into.
| Limit | Claude | OpenAI | Gemini |
|---|---|---|---|
| Discount | 50% | 50% | 50% |
| Max per batch | 100,000 requests or 256 MB | 50,000 requests, 200 MB file | 2 GB input file |
| Window | 24 hours, most under 1 hour | 24h, fixed | Target 24 hours, expires at 48 |
| Input format | Requests array | JSONL file | Inline under 20 MB, or JSONL |
| Results kept | 29 days | 30 days | 6 weeks |
Figures come from Anthropic’s batch processing documentation, OpenAI’s Batch guide and Google’s Gemini Batch API page, read in October 2026.
A 600-page site fits in one batch on every provider. The limits start to matter when you batch per-keyword FAQ generation across several sites.
What Does a Batch Actually Cost Next to a Live Call?
Anthropic lists batch prices per model. Claude Sonnet 5.5 is $1 per million input tokens and $5 per million output tokens, and Haiku 4.5 is $0.50 and $2.50.
Take 2,000 meta descriptions at roughly 600 input and 120 output tokens each. That is 1.2M input and 0.24M output tokens.
On Sonnet 5.5 batch pricing that is $1.20 plus $1.20, or $2.40 in total. Check the arithmetic against your own token counts before you budget from it.
OpenAI and Google state the 50% rule but price per model, so look up your model’s batch rate rather than halving a number from memory.
Why Do Batch Results Come Back in the Wrong Order?
Because requests are processed independently and concurrently. Anthropic’s documentation says results can be returned in any order, and OpenAI warns that output line order may not match input order.
Every request therefore needs an ID you can join on. Claude requires a custom_id of 1 to 64 characters matching ^[a-zA-Z0-9_-]{1,64}$.
Use the WordPress post ID plus the field name, such as p4821-meta. A slug with a dot or a space is rejected.
Never write results back by line position. One errored request shifts every row below it, and you publish a meta description onto the wrong post.
How Do You Handle Errored and Expired Requests?
Claude reports four outcomes per request: succeeded, errored, canceled and expired. Errored, canceled and expired requests are not billed, per Anthropic’s documentation.
OpenAI sends failures to a separate error file named by error_file_id, and expired requests carry the batch_expired code. Gemini ends a job in one of JOB_STATE_SUCCEEDED, FAILED, CANCELLED or EXPIRED.
Loop over the output, collect every ID that did not succeed, and resubmit only those as a smaller batch. Do not rerun the whole file.
Store the ID-to-post map before you submit. If the batch expires you need it to rebuild the retry set without re-reading the site.
Does Prompt Caching Still Work Inside a Batch?
On Claude, yes, but only on a best-effort basis. Anthropic notes that batch requests run concurrently and in any order, so cache hits are not guaranteed.
Because a batch can take longer than five minutes, the documentation suggests the 1-hour cache duration when requests share a long prefix such as a style guide. Gemini says context caching is supported at standard caching rates.
Our comparison of prompt caching across Claude, OpenAI and Gemini covers the prefix rules in more detail.
Batch results can be returned in any order, and may not match the ordering of requests when the batch was created.
– Per Anthropic’s published batch processing documentation
Use a batch API for any SEO generation job that can wait a day. The 50% discount is the same everywhere, so choose on limits, retention and the model you already trust.
Key every request with a stable ID, retry only the failures, and download results before the retention window closes.
Frequently Asked Questions
How much cheaper is a batch API than a live call?
All three providers state a 50% discount. OpenAI’s Batch guide and Google’s Gemini Batch API page both use the 50% figure.
How long does a batch take?
Anthropic’s documentation says most batches finish within 1 hour and expire after 24. OpenAI uses a fixed 24h window. Gemini targets 24 hours and expires jobs at 48.
Can I match results to inputs by position?
No. Both Claude and OpenAI say output order may differ from input order. Join on custom_id instead.
Am I billed for requests that fail or expire?
On Claude, errored, canceled and expired requests are not billed, per Anthropic’s batch documentation. Check the billing terms of the other two before relying on the same behavior.
How long can I download results?
Claude keeps results for 29 days after creation and OpenAI deletes output files 30 days after completion. Gemini stores results for 6 weeks.
DesignCopy Editorial Team
English 
한국어 