Which LLM providers offer both batch API and prompt caching discounts?

Anthropic, Google, and OpenAI all offer both batch processing and prompt caching. Anthropic's Claude offers 50% batch discounts plus 90% cache reads; OpenAI provides 50% batch discounts and 50-90% cache discounts depending on duration; Google's Gemini offers similar dual-discount structures. AWS Bedrock and Azure OpenAI inherit these features from their underlying models.

How long does prompt cache stay active before expiring?

Prompt cache duration varies by provider: Anthropic's cache lasts 5 minutes of inactivity, OpenAI's ephemeral cache persists for the session only, and Google Vertex AI maintains cache for up to 1 hour. Cache expiration means you'll pay full input prices again on the next request if you exceed the time window between API calls.

What's the minimum request size required to use batch API discounts?

Most providers have no strict minimum batch size, but OpenAI recommends batches of at least 10-100 requests for worthwhile setup overhead. The real constraint is timing: batch jobs typically complete within 24 hours, so you need workloads that tolerate 30-minute to 24-hour delays rather than requiring real-time responses for individual requests.

Do batch API discounts apply to fine-tuned models?

Yes, batch API discounts apply to fine-tuned models with all major providers. OpenAI offers 50% batch discounts on fine-tuned GPT-4 and GPT-3.5 models; Anthropic supports batch processing for fine-tuned Claude models at the same 50% discount. The batch discount stacks on top of your base fine-tuned model pricing, not the standard model rates.

METHODOLOGY DEEP-DIVE · VERIFIED 15 JUL 2026

What are batch API and prompt caching discounts for LLMs?

Batch API and prompt caching are two discount mechanisms offered by LLM providers that can reduce costs by 50% and 90% respectively. Batch APIs deliver ~50% savings by processing requests asynchronously with delayed responses, while prompt caching cuts costs by ~90% on repeated context portions by storing them temporarily. Both discounts are stackable with most providers, enabling combined savings exceeding 95% on applicable workloads.

Every price on this site — the calculator, the full table, every worked example in our use-case guides — is the standard, pay-as-you-go list rate. We say this explicitly and often, because two features available from every major provider we track can meaningfully lower a real production bill below those numbers, and most people never turn them on.

Batch processing: roughly 50% off, if you can wait

Every major provider offers a batch API: submit a large set of requests together, get results back within a defined turnaround window — typically same-day or next-day rather than instantly — at roughly half the standard price. The trade-off is exactly what it sounds like: no real-time response. This is the single biggest lever available for any workload that doesn't need an immediate answer.

Where batch processing genuinely fits

Overnight data processing, reports generated for the next morning

Bulk content generation queued for later publishing

Large-scale classification or tagging jobs run on a schedule

Any workload currently running synchronously that doesn't need to be

Look back at our high-volume classification worked example: a $28.50/month workload on the cheapest tier drops to roughly $14/month with batch pricing alone, before you've changed a single model choice. For any workload that's currently synchronous purely out of habit rather than genuine need, this is worth checking before anything else.

Prompt caching: roughly 90% off the repeated portion

Prompt caching targets a different pattern: when the same block of text — a system prompt, a set of instructions, a reference document, a codebase — gets sent on every call, with only a small portion actually changing between requests. Providers can cache that repeated portion and charge roughly 90% less for it on cache hits, while the changing portion is billed at the normal rate.

Where prompt caching genuinely fits

Coding assistants re-sending the same codebase context on every call

Support bots referencing a fixed knowledge base on every conversation

Any system prompt with detailed instructions, reused across thousands of calls

Multi-turn conversations where earlier turns stay in context

This is precisely why our coding assistants guide calls caching a bigger lever than model choice — a coding assistant's input is dominated by the same codebase context repeated call after call, which is exactly the shape caching is built for.

Can you use both at once?

Where a provider's API supports it, yes — batch and caching address different parts of the cost structure (turnaround time versus repeated content) and generally aren't mutually exclusive. A large, non-urgent job with a heavily repeated system prompt is the best-case scenario for stacking both discounts. Check your specific provider's current documentation for exact interaction rules, since implementation details vary and change.

The honest reason we don't bake this into every number on the site

Batch turnaround windows and cache hit rates vary by provider and by how consistently your actual traffic repeats the same content — a number we can't know without seeing your specific workload. Publishing a single "with discounts" figure would imply a precision we don't actually have. The list prices we show are the honest baseline; these two levers are what you check next, specific to your own usage pattern.

General availability and approximate discount levels verified across major providers' documentation, 15 July 2026. Exact terms, turnaround windows, and cache TTLs vary by provider — check current documentation before committing a production workload to either.