OpenAI
Check the pricing page for cached-input, Batch API, priority, and flex rows, then use the prompt caching guide to confirm eligibility and cached-token reporting.
Pricing guide
LLM API cache and batch pricing: when headline token prices are not enough
Input and output token prices are still the baseline. But repeated context, delayed jobs, and optional processing tiers can move the real cost check to cached input, cache write, batch, priority, or flex rows.
Decision path
Start with the workload timing
A lower headline token price is not always the cheapest production shape. First decide whether the workload is interactive, repeated, offline, or tied to a processing tier.
| Workload pattern | Pricing rows to inspect | Best next step |
|---|---|---|
| One-at-a-time product requests | Input and output | Estimate the monthly request shape in the calculator. |
| Repeated instructions, documents, or system context | Cached input and cache write | Start from the cache reuse preset, then change the hit rate. |
| Offline analysis or large jobs that can wait | Batch input and batch output | Estimate the workload with batch pricing, then compare which models expose batch rows. |
| Latency or capacity tier decisions | Priority and flex input/output | Compare the optional processing rows beside standard prices. |
Calculator
Use the cache reuse preset when repeated context is plausible
The calculator includes a cache reuse preset with 8k input tokens, 1k output tokens, 5k monthly requests, and a 50% cache hit rate. Treat that as a starting point, not a promise from any provider.
If a selected model exposes cached-input pricing, the calculator uses that row for the cached share. If it does not, the estimate falls back to ordinary input pricing so the scenario still has a conservative first-pass cost.
Cache check
- Does the repeated text meet the provider's caching requirements?
- Is cache write pricing separate from cache hit pricing?
- Does the cache expire before the repeated requests arrive?
Pricing data sourced from our catalog. Check data sources for provenance and freshness.
Calculator + compare
Estimate cache and batch scenarios, then inspect the underlying pricing rows
The calculator can estimate cached-input reuse and use batch input/output prices when those rows are available. The compare page shows the underlying cached input, cache write, batch input/output, priority, and flex rows side by side so you can see which pricing dimensions each route actually exposes.
Use the calculator for the monthly workload estimate and Compare for the row-level check. This is especially useful when two models look similar on standard input/output cost but differ in optional pricing rows.
Do not skip
Batch and cache rows do not replace provider documentation. Confirm eligibility, rate limits, retention, latency, and billing behavior before treating a route as production-ready.
Provider docs
Check the provider rule before trusting the estimate
Anthropic
Check cache write, cache hit, batch, and long-context notes before treating a Claude workload as a simple input/output estimate.
Gemini
Check context caching token prices, storage prices, and cache TTL behavior before estimating repeated-context workloads.
Limits
What this guide does not decide
This guide does not rank model quality, latency, quotas, regional availability, tool charges, account discounts, or final invoices. It also does not guarantee cache eligibility for any prompt. Use it to choose the right pricing rows, then verify the production contract with the provider.