Open-weight models on Bedrock can now cache the prompt prefix you choose
Amazon Bedrock now lets developers mark exactly which part of a prompt to cache for open-weight models, instead of relying on best-effort automatic reuse.

Amazon says open-weight models on Amazon Bedrock now support explicit prompt caching. Developers mark the reusable part of a prompt, and later requests read it from cache, which cuts input cost and latency, according to an AWS Machine Learning blog post.
One request option and a breakpoint marker
Bedrock already offered implicit caching for these models. It needs no changes to a request, and Bedrock picks eligible prefixes on a best-effort basis. The new mode gives developers control over what is stored.

On the Chat Completions API, a request sets prompt_cache_options to mode: explicit and adds one or more prompt_cache_breakpoint markers on text content parts. Everything up to and including a marker is cached. Anything after the last marker is processed on every request and billed at the standard input rate. The OpenAI Python SDK sends the option through extra_body.
The post stresses that the two pieces only work as a pair. If a request enables explicit mode but includes no marker, nothing is stored. If it includes a marker but leaves explicit mode off, Bedrock falls back to implicit behavior.

Limits depend on the model. The post points readers to each model card for the minimum prefix length, the maximum number of breakpoints, the cache lifetime (TTL) and the supported ttl values. The TTL resets with each cache hit. The post does not name specific open-weight models or Regions, and says only that availability and caching support vary by Region.
Each response reports cache activity in usage.prompt_tokens_details, through cached_tokens and cache_write_tokens. Across a workload, AWS suggests tracking the CacheReadInputTokenCount and CacheWriteInputTokenCount runtime metrics.
Why the write bill is the point
Cache reads cost less than standard input. Cache writes can cost more, depending on the model. AWS notes that implicit caching can also write content that is never reused, such as a customer question that differs every time. Explicit caching limits writes to the shared prefix, so the higher write rate is not paid on content nobody reads back.
AWS demonstrates this with a support assistant that holds a roughly 4,000-token product manual in the system prompt. In implicit mode, the first call writes the prompt. Calls 2 and 3 read the manual from cache but each write a few extra tokens, because the cached prefix grows to cover the new question. With explicit caching, only the first call writes, and later calls read the prefix and write nothing. The post says all this is for a first run, and that repeating the run within the TTL gives cache reads on every call in implicit mode too.
The same post tests implicit caching with the OpenCode coding agent, version 1.18.35, pointed at the Bedrock endpoint as an OpenAI-compatible provider. Each agent step resends the system prompt, tool definitions and history, so later requests read almost everything except the new tail from cache. AWS notes that OpenCode does not record cache_write_tokens for this kind of provider, so its write counter stays at zero.
A pattern Bedrock already uses for other models
Technical significance: the mechanism is the same one AWS documents for other families on Bedrock, which makes the open-weight version easier to place. According to the Bedrock user guide, Claude models accept up to four cache checkpoints with a 5-minute or 1-hour TTL, and minimum prefix sizes range from 512 to 4,096 tokens depending on the model. For OpenAI’s GPT-5.6 and later models on Bedrock, the guide lists the same prompt_cache_breakpoint field, a 1,024-token minimum, up to four cache writes per request, a 30-minute TTL, and cache writes billed at 1.25 times the uncached input rate.
Those numbers come from other model families, so they should not be read as the open-weight limits. The open-weight models have their own model cards, and the blog post gives no figures for them. The guide also says that caching is not supported with batch inference, that it applies to on-demand endpoints, and that a cache entry lives on the server that handled the request. As a result, requests routed to another Region, or sent before an identical request has started responding, write a new entry instead of reading one. AWS adds that cross-Region inference can mean extra cache writes at times of high demand.
For agent builders this matters because the savings depend on prefix discipline. AWS advises putting static content first, keeping it identical between requests, and moving timestamps and request IDs to the end. It also notes that if an early part of the prompt changes, the cache is rebuilt from that point.
What the announcement leaves open
The post gives no discount percentage, no latency figures and no list of supported models. Savings and speedups are described in general terms, and the usage fields in each response are offered as the way to measure them on real traffic. Prices are on the Bedrock pricing page. AWS also repeats that, as with other Bedrock models, prompts and outputs are not shared with the model provider or used to train models.
Developers who want to try it need access to a supported open-weight model, IAM permission to invoke it, and the openai and aws-bedrock-token-generator Python packages. The example code authenticates with a short-term Bedrock API key valid for up to 12 hours.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.
You may also like

Claude on Amazon Bedrock can now be forced to return valid JSON

Codex can now run on Amazon Bedrock, with AWS logins instead of OpenAI keys

ChatGPT desktop app can now run on Bedrock inside AWS GovCloud, with OpenAI sites blocked
