AWS’s November 16, 2026 guide to Amazon Bedrock prompt caching explains how teams running document assistants or coding agents can reuse repeated context. Caching saves money when cheaper reads outweigh cache-write charges—not simply because a prompt is long.
For Kimi K3.1 Pro Highspeed at the rates below, one write followed by one successful read costs 35% less for the repeated text than two uncached inputs. That is an illustrative calculation, not a measured or guaranteed reduction in the whole bill.
A prefix is the unchanged beginning of a request, such as instructions followed by a reference document. Prompt caching reuses earlier processing of that text; it is not permanent assistant memory.
When does a cache write pay for itself?
Use one specific price plan: Kimi K3.1 Pro Highspeed, Standard tier, Global cross-Region inference. AWS’s list prices, checked on November 16, 2026, are US$4 per million ordinary input tokens, $5 per million cache-write tokens, $0.20 per million cache-read tokens and $20 per million output tokens. Tokens are the units used to count text processed by the model.
Assume each request contains the same eligible 10,000-token prefix. A successful read reuses all of it; a miss writes all of it again. The comparison below prices that prefix as ordinary input on every request. It excludes changing input, generated answers, tools, retrieval, retries, infrastructure, tax, discounts and implementation work.
Use of the 10,000-token prefix | Cache charges | All-uncached comparison | Difference |
|---|---|---|---|
One write, no reuse | $0.050 | $0.040 | 25% more |
One write, one successful read | $0.052 | $0.080 | 35% less |
Two requests, both write | $0.100 | $0.080 | 25% more |
One write, nine successful reads | $0.068 | $0.400 | 83% less |
The arithmetic for two requests is 10,000 ÷ 1,000,000 × ($5 + $0.20) = $0.052. Without caching, it is 10,000 ÷ 1,000,000 × ($4 + $4) = $0.080. The first successful full-prefix read therefore repays the extra cost of the initial write.
A cache read is 95% cheaper than ordinary input at these rates, but that does not make the first two requests 95% cheaper: their combined prefix saving is 35%. If neither request reads the cache, the prefix costs more instead.
For a mixed workload at these rates, the break-even read share is ($5 − $4) ÷ ($5 − $0.20), or about 20.83%. Savings require a higher share of the combined cache-read and cache-write token volume to be billed as reads. Count tokens, not requests: a short hit does not offset a much larger rewrite. This threshold applies to the modeled prefix charges, not every Bedrock model or the account’s entire bill.
Compare explicit caching with your current application, not just a no-cache estimate. Kimi supports automatic, or implicit, caching by default. An existing application may already receive discounted reads. The table compares caching with hypothetical fully uncached input; it does not measure the additional saving from switching to explicit controls.
Which API and location do these figures apply to?
The Kimi K3.1 Pro Highspeed model card lists these conditions:
Explicit caching works through Responses, Chat Completions and InvokeModel, but not through Converse for this model.
The minimum prefix at a cache checkpoint is 256 tokens. The listed cache retention is at least 30 minutes; neither condition guarantees a hit.
The model is available through US, India and Global cross-Region inference. The geographic profiles restrict routing to their respective geographies; Global can route worldwide to supported commercial AWS Regions.
The US and India Standard profiles have different list prices: $4.40 ordinary input, $5.50 cache write, $0.22 cache read and $22 output per million tokens. Use the profile your data-location requirements permit, then calculate its costs. Do not apply the cheaper Global tariff to a geographically restricted request.
What should go before the cache breakpoint?
Keep reusable instructions and reference material first, with changing questions, timestamps and request identifiers afterwards. In the tutorial’s Chat Completions pattern, explicit caching requires both prompt_cache_options with mode explicit and a prompt_cache_breakpoint marker. Put the marker after the reusable prefix. Content after the last marker is billed as ordinary input, avoiding cache-write premiums for text that will not be reused.
A correct breakpoint still does not guarantee a cache hit. AWS’s caching documentation says a changed prefix, an expired entry or a request sent to a different serving server can prevent reuse. A request arriving before an identical request starts responding can also write a new entry. Successful hits refresh the retention period, while cross-Region routing can increase writes. Include idle gaps and concurrent traffic when assessing costs.
How do you check the actual bill?
In the tutorial’s Chat Completions responses, usage.prompt_tokens_details.cached_tokens counts reads and cache_write_tokens counts writes. Subtract both from usage.prompt_tokens to find ordinary input. Do not charge cached tokens again as ordinary input.
For the Global Standard example, the token charge in dollars is:
(ordinary input tokens × 4 + cache-write tokens × 5 + cache-read tokens × 0.20 + output tokens × 20) ÷ 1,000,000.
Track both CacheReadInputTokenCount and CacheWriteInputTokenCount in Bedrock’s CloudWatch runtime metrics. Compare a representative period with the application’s existing caching behavior, using the same model, profile and service tier. Read counts alone cannot establish net savings.
Include output and request size when comparing alternatives, too. The separate Muse Ember guide to long-request costs illustrates why the price of one input category is not a complete task budget; its tariffs are not the Bedrock rates used here.
Under the two-request assumptions above, 1,000 independent write-and-read pairs save $28 on prefixes: $52 instead of $80. That is a volume calculation, not a customer result. Compare the potential saving with engineering effort and the complete bill, while checking that answers remain useful.
If the larger decision is whether to operate the model yourself, the Kimi K3.1 Pro self-hosting hardware and licence guide addresses that separate deployment choice. For an existing Bedrock application, start with its actual cache writes and reads: frequent reuse can justify caching; one-off or repeatedly rewritten context may not.
