Batch API pricing: the 50% discount and its catch
The big three providers all sell asynchronous batch processing at half the standard price, in exchange for waiting up to a day. This page compares their rules, sizes and limits, and works out what the discount saves on a real workload. Model your own volume in the cost calculator, which has a batch option; the wider topic is AI costs and pricing.
Prices last verified against official provider pages on . How we verify prices. Written by the Chelure Enterprises Editorial Team.
How the providers compare
| Anthropic (Message Batches) | OpenAI (Batch API) | Google (Gemini Batch API) | |
|---|---|---|---|
| Discount | 50% off input and output | 50% off synchronous prices | 50% of the standard interactive price |
| Turnaround | Most batches finish within 1 hour; expire if not complete within 24 hours | Complete within 24 hours, often more quickly (only 24h supported) | Target 24 hours, usually much quicker; jobs pending or running over 48 hours expire |
| Batch size | 100,000 requests or 256 MB, whichever first | 50,000 requests; input file up to 200 MB | Input file up to 2 GB; inline requests under 20 MB |
| Results kept | 29 days | Output and error files expire after 30 days by default | Not stated on the page we read |
| Caching | Discounts combine; 1-hour cache suggested | Not covered on the batch page | Batch cached-input price same as standard for Gemini 3.1 Pro; storage fee applies |
| Rate limits | Own limits on batch requests | Separate pool, does not use per-model limits | See Google's rate-limits page |
What it saves
Take 1,000 requests a day for 30 days at 1,500 input and 400 output tokens, on Claude Sonnet 5.5 ($2 input, $10 output per million tokens), or on GPT-6.1 Sol, which has the same list prices. Standard cost is $210.00 a month and batch cost is $105.00, a saving of $105.00. For jobs you run once, such as classifying an archive, the saving is the same 50 percent of a larger number.
Anthropic says its batch and prompt caching discounts can be combined, and its pricing page points to the caching section for how the multipliers interact. Claude Sonnet 5.5 cache reads are $0.10 per million tokens at standard pricing. See the prompt caching guide for the cache multipliers, and confirm the combined rate on Anthropic's pricing page before budgeting.
When not to use batch
- Interactive products: a user waiting for an answer cannot wait hours.
- Time-sensitive pipelines: batches that miss their window expire; Google's expire after 48 hours and Anthropic's after 24.
- Features that are not offered in batch: Anthropic notes, for example, that its fast mode is not available with the Batch API.
Other providers may offer batch discounts too; we only list those we verified on their own pages. Compare list prices in the price comparison.
Frequently asked questions
How much cheaper is a batch API?
Anthropic, OpenAI and Google each list 50 percent off the standard price for their batch services. Anthropic says all usage in the Message Batches API is charged at 50 percent of standard prices, OpenAI says batch costs 50 percent less than synchronous APIs, and Google prices its Batch API at 50 percent of the standard interactive cost.
What is the catch?
You do not get an answer immediately. OpenAI's batches complete within 24 hours (often more quickly), Anthropic's most often finish within an hour but expire if not complete within 24 hours, and Google's target turnaround is 24 hours, with jobs pending or running for more than 48 hours expiring.
Can I combine batch and prompt caching?
On Anthropic, yes: its pricing page says batch and caching discounts can be combined, and it suggests the 1-hour cache for batches because they can take longer than 5 minutes. On Google, the Gemini 3.1 Pro batch table lists cached input at the same price as standard, so caching adds no extra discount there.
How big can a batch be?
Anthropic: up to 100,000 requests or 256 MB. OpenAI: up to 50,000 requests and a 200 MB input file. Google: input files up to 2 GB, inline requests under 20 MB, with no maximum request count stated.
Is batch good for chatbots?
No. Batch suits work that can wait: evaluations, classification of large archives, embeddings, content generation and nightly reports. Anything a user is waiting on should use the standard API.
Sources
- OpenAI: Batch API guide
- Anthropic: batch processing
- Anthropic: pricing (batch, stacking)
- Google: Gemini Batch API
- Google: Gemini API pricing
Primary sources opened on November 13, 2026. Prices come from our verified price table (see the methodology).