Home › Tools › Prompt caching calculator

LLM Prompt Caching Cost Calculator

Prompt caching is the single largest cost lever in modern LLM APIs. On a 10,000-token system prompt at 80% cache hit rate, the monthly bill drops by 60-78% depending on the model. This page shows the cached vs. uncached cost for all 28 models in the dataset.

Cached vs. uncached cost per model

Cost of a 10,000-token prefix sent 1,000 times a month, at 80% cache hit rate, including write premium. Sorted by saving.

Uncached = 10,000 tokens x 1,000 requests at full input rate. Cached = 80% reads at cached rate + 20% uncached + writes at 1.25x (12 writes per month for a 5-minute TTL). Output tokens are excluded — caching discounts input only.

How prompt caching pricing works

Every provider bills cached prompts in three token classes:

• Write premium. Storing a prefix costs more than sending it once — 1.25x on Anthropic (including Bedrock), storage-based on Google.
• Read discount. A cache hit costs 0.10x base input on Anthropic, $0.075 on Gemini Flash — a 90% saving per token read.
• Write frequency. The cache expires. Every refresh pays the premium again. At steady traffic this is usually the largest line item.

When caching loses money

• Prefixes that change every call. User-specific instructions at the top guarantee a write and zero reads.
• Long gaps between requests. A 20-minute interval means multiple rebuilds per reuse.
• Short prompts below the minimum. Nothing is cached, and you pay full rate.
• Output-heavy workloads. Caching does nothing for output, which is 4-5x input.

Keep going

• All 28 models ranked by price
• Head-to-head comparisons
• Prompt Caching Calculator: Bedrock vs Gemini (blog)
• Bedrock Prompt Caching Pricing Guide (blog)
• Context Caching on Gemini (blog)

Calculate a real monthly bill

FAQ

What is prompt caching?

Prompt caching stores a repeated prefix (system instructions, tool schemas, retrieved documents) so the model does not re-process it on every request. Cached tokens are billed at a deep discount — 0.10x on Anthropic, $0.075 on Gemini Flash.

How much can prompt caching save?

On a typical workload with a 40,000-token prefix and 80% cache hit rate, prompt caching reduces the input bill by 60-78%. The exact saving depends on prefix size, traffic, and cache policy.

Does Bedrock support prompt caching?

Yes. Bedrock bills the same token classes as the Anthropic API: 1.25x to write, 0.10x to read, with a 5-minute TTL refreshed by use. Up to four cache breakpoints are supported.

Is Gemini context caching cheaper than Bedrock?

For irregular traffic, yes — Gemini implicit caching has no write premium. For steady high-volume workloads, Bedrock can win if writes are kept infrequent with a longer TTL.

What is the minimum cacheable length?

Roughly 1,024 tokens for Sonnet and Opus class models, 2,048 for Haiku. Shorter prompts are silently not cached.

Keep going

• Tokens to cost
• All models
• Comparisons
• All 28 models ranked by price
• Run your own volumes in the calculator

Calculate a real monthly bill