Bedrock prompt caching looks like a 90% discount — then the invoice shows a 15% saving. The gap is not a rounding error. On Amazon Bedrock, caching is billed in three separate token classes, and only one of them gets the advertised discount. If you model the others as free, your budget will be wrong by hundreds of dollars a month.
This guide explains Bedrock prompt caching pricing the way the invoice actually works: what you pay to write a cache, what you pay to read it, how the time-to-live resets, and when the same workload is cheaper without caching at all. All numbers below use the October 2026 rate card and can be reproduced in the AI cost calculator on rabayid.com.
The three token classes that decide your bill
Ignore any single percentage. A cached prompt on Bedrock has exactly three costs:
- Write premium. Storing a prefix costs 1.25× the base input rate. This is charged once per prefix per TTL window.
- Read discount. A cache hit costs 0.10× the base input rate — a 90% saving per token read.
- Write frequency. The cache expires. Every refresh pays the premium again. At steady traffic, this is usually your largest line item.
If your model only includes the read discount, it will always under-estimate the bill. The write count is the missing term.
Bedrock prompt caching rates for Claude (2026)
Bedrock mirrors Anthropic's token classes. What changes is where the cache lives, how it is scoped, and which models support it. Here are the current list prices per 1M tokens:
| Model | Base input | Cache write | Cache read | Output |
|---|
| Claude Opus 5 | $5.00 | $6.25 | $0.50 | $25.00 |
| Claude Sonnet 5 | $2.00 | $2.50 | $0.20 | $10.00 |
| Claude Haiku 4.5 | $1.00 | $1.25 | $0.10 | $5.00 |
Two rules follow directly from this table:
- The write premium is always 1.25× base input. A write costs more than sending the prefix once. That only makes sense if the prefix is read again inside the TTL.
- The read discount is always 0.10×. Each read saves 90% relative to uncached input.
There is no storage line. You are not billed per hour for keeping the five-minute cache — the premium covers it.
Worked monthly math: 36,000 requests
Take a support assistant running on Claude Sonnet 5: a 40,000-token prefix (system prompt, tool schemas, policy corpus), 2,000 tokens of conversation per request, 500 output tokens, 1,200 requests a day for 30 days — 36,000 requests a month.
| Line item | No caching | 5-minute TTL |
|---|
| Uncached input | $3,024 | $144 |
| Cache writes | — | $864 |
| Cache reads | — | $288 |
| Output | $180 | $180 |
| Total | $3,204 | $1,476 |
Cache writes are 58% of the cached bill even though the prefix never changes. The cache is rebuilt 8,640 times a month because traffic is steady but the TTL is short. That is the whole story of Bedrock prompt caching pricing: the discount is real, and the write frequency decides whether you keep it.
TTL, breakpoints and minimums: the fine print that bills you
Five Bedrock facts change the math more than any discount:
- Five-minute TTL, refreshed by use. Active traffic keeps the prefix warm. A nightly batch every four hours writes far more than it reads.
- Up to four cache breakpoints. Place them at the end of blocks that actually repeat: system instructions, tool schemas, then retrieved documents. A breakpoint in the middle of a changing conversation caches noise.
- Minimum cacheable length. Roughly 1,024 tokens for Sonnet and Opus, 2,048 for Haiku. Shorter prompts are silently not cached and bill at full rate.
- Cache is per-model and per-region. Switching region or model invalidates the prefix and forces a fresh write.
- Longer TTL costs more per write. Anthropic prices a one-hour cache at 2× base input on the API. If available in your Bedrock region, it can still win because it writes 12× less often — confirm availability before budgeting on it.
Break-even: when does one write pay for itself?
The monthly model is short enough to keep in a spreadsheet cell:
monthly = N x (P x r_read + V x r_in + O x r_out)
+ W x P x r_write
N = requests per month P = cached prefix tokens
V = variable input O = output tokens
W = writes per month r_* = rate per token
On Sonnet 5, a write costs $0.50 per million tokens above base; each read saves $1.80 per million. Caching pays for itself on the first re-read inside the TTL. The question is never whether to cache — it is how often you rebuild.
Where Bedrock pricing differs from the direct Anthropic API
The token classes match, but operations do not. Four differences matter for cost:
- Scoping. On Bedrock, caches are scoped to your AWS account and region. Multi-region workloads pay the write premium in every region.
- Identity. IAM roles and cross-account inference profiles can split traffic across several caches instead of one.
- Model support. Not every Bedrock model supports caching. If a model falls back to an uncached path, you pay full input rate with no warning in the response.
- Agreement pricing. Committed-use discounts apply to base tokens, not necessarily to cache writes. Verify how your enterprise agreement treats them.
None of these appear on a price sheet. All of them appear on the invoice.
Use a prompt caching calculator bedrock workflow before you enable it
The safest workflow is boring, and it works:
- Open the AI model calculator and select Claude Sonnet 5 from the dataset.
- Enter your real prefix size — not a guess. Read it from API usage logs.
- Model three scenarios: 0% cached, your measured hit rate, and 100% cached. The spread is your risk.
- Compare models side by side in the head-to-head page before committing.
- Enable caching on one route, measure cached token counts for a week, then decide.
Do not enable caching everywhere and hope. Measure the write line first.
When caching loses money on Bedrock
- 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 with a false sense of optimisation.
- Output-heavy workloads. Caching does nothing for output, which is 4–5× input on every Claude model above.
FAQ
How much does Bedrock prompt caching cost?
You pay 1.25× base input to write a prefix, 0.10× to read it, and nothing extra for the five-minute cache. On Sonnet 5 that is $2.50 per million written tokens and $0.20 per million read tokens.
Is the write premium worth it?
Yes, if the prefix is re-read inside the TTL. One re-read on Sonnet 5 more than covers one write. Without re-reads, you pay a 25% premium for nothing.
Can I use a one-hour cache on Bedrock?
Anthropic prices it at 2× base input on the direct API. Availability varies by region and model on Bedrock, so confirm in your console before budgeting on it.
Why is my hit rate high but my bill high?
Because writes dominate. Steady traffic with a short TTL rebuilds the cache thousands of times. Check the write count, not only the hit rate.
Related reading
Your next step is one number: how many times per month your cache is rebuilt. If it is more than your request count divided by ten, the TTL costs more than the model. Run your prefix through the free AI cost calculator and read the write line yourself.
