Batch API Discount: When 50% Off Is Worth It
Batch APIs let you submit many requests at once and get results later, usually within hours. In return, providers charge about half the standard price.
Examples of batch prices
- Claude Sonnet 5.5: $1 input / $5 output per million tokens, versus $2 / $10
- GPT-6 Sol: $1 input / $5 output, versus $2 / $10
- Gemini: batch processing is offered at 50% off standard rates
Good batch candidates
- Classifying or tagging large datasets
- Summarizing archives of documents
- Generating product descriptions or translations in bulk
- Nightly or weekly reports
- Running evaluations on a model
Poor candidates
Anything a user is waiting for in real time, such as chat, live support or interactive tools. Batch results arrive on the provider's schedule.
Combine with other savings
Batch discounts can stack with a cheaper model. A job moved from a top model to a small one and then to batch can cost a small fraction of the original. Cache rules vary, so check the provider's documentation before assuming discounts add up.
How batch processing works
Instead of sending requests one by one and waiting for each reply, you submit a group of requests together. The provider processes them when capacity allows, usually within hours and often faster, and you collect the results afterwards. Because the provider can fill spare capacity, it charges about half the standard rate.
What the savings look like
A nightly job processes 10 million input tokens and generates 2 million output tokens on Claude Sonnet 5.5. At standard rates that is about $40 a night. Through batch, at $1 / $5 per million, it is about $20 a night, a saving of about $600 over a 30-night month.
Now combine batch with a smaller model. The same job on Claude Haiku 5.5 at half price, $0.05 / $0.25 per million, costs about $1 a night. Provided Haiku is good enough for the task, the monthly bill falls from about $1,200 to about $30.
When batch fits
- The result is not needed immediately
- Work arrives in large groups, such as imports or archives
- Jobs can be retried safely
- You run scheduled reports or evaluations
How to design a batch workflow
1. Collect requests into a file or list with an ID for each.
2. Submit the batch and store the batch ID.
3. Poll for completion on a schedule, rather than waiting in a loop.
4. Download results and match them to inputs by ID.
5. Retry only the failed items.
Expect that a small share of requests may fail or expire, so build in a retry step.
Things to check in the documentation
- Maximum requests and file size per batch
- The completion window, often up to 24 hours
- Whether the discount applies to cached input and long prompts
- Separate rate limits for batch jobs
- How long results stay available for download
When not to use batch
Anything interactive, such as chat, live support, autocomplete or voice, needs fast responses. Do not move user-facing traffic to batch just for the discount. A middle path is to batch background work, like generating summaries after a call ends, while keeping live replies on the standard API.
Frequently asked questions
Is the batch discount the same everywhere?
Roughly half off is common across the major providers, but check each one, as terms differ.
How long do batch jobs take?
Most finish within hours, but providers usually promise completion within 24 hours.
Can I cancel a batch?
Most providers allow cancellation, and you pay only for completed work. Confirm in the documentation. Learn more about other ways to cut costs.
Estimate the effect
Run your volume through the cost calculator and halve the result for the batch scenario. More ideas: 7 ways to cut your LLM bill.
Rates checked November 4, 2026.
Track spend by team and project: the AI Spend Tracker spreadsheet is launching soon.