Amazon Bedrockで OpenAI GPT-5.6 モデル向け明示的プロンプトキャッシュが利用可能にIntroducing explicit prompt caching for OpenAI GPT-5.6 models on Amazon Bedrock

Amazon Web Services has made OpenAI's GPT-5.6 family generally available on Amazon Bedrock, adding explicit prompt caching that lets developers specify exactly which portions of a prompt should be cached and reused. The change matters because inference cost and latency remain among the biggest practical obstacles to running large language models at scale, and finer control over caching targets both directly.

The release covers three GPT-5.6 variants, Sol, Terra, and Luna, now offered through Bedrock's managed API alongside the caching feature. Bedrock is AWS's fully managed service for accessing foundation models from multiple providers through a single interface, so the addition extends the catalog developers can call without provisioning their own infrastructure or managing model hosting themselves.

Prompt caching addresses a common inefficiency in LLM applications. Many requests share large, unchanging blocks of text, including system instructions, tool definitions, retrieved documents, or few-shot examples, that the model would otherwise reprocess on every call. Caching stores the intermediate computation for those repeated segments so subsequent requests can reuse it, reducing the tokens that must be processed and, in turn, the cost and response time.

The distinguishing detail here is that the caching is explicit. With implicit approaches, the platform automatically detects and reuses repeated prefixes. Explicit caching instead gives developers precise control over which parts of a prompt are marked for caching, letting them place cache boundaries deliberately around the stable portions of a prompt while leaving dynamic sections uncached. That control appears aimed at workloads with predictable structure, such as agents that reuse the same instructions and tool schemas across many turns.

For teams already running GPT workloads, AWS frames the update partly as a migration path. The accompanying guidance covers getting started, configuring explicit caching, and moving existing GPT-based applications onto Bedrock to reduce inference cost. Because the models are served through Bedrock's standard interface, applications that call OpenAI models elsewhere can in principle be redirected with adjustments to endpoints and request formatting, though the actual effort will depend on how each application is built.

The feature fits a broader industry pattern. Prompt caching has become a standard cost-optimization lever across major model providers, and several vendors already offer some form of it on their own platforms. Bringing explicit caching to OpenAI models on Bedrock aligns them with caching capabilities available for other model families on the service, giving developers a more consistent set of tools regardless of which provider they choose.

The agent framework context is also relevant, since caching is particularly useful there. Agentic applications tend to issue many sequential model calls that repeat the same system prompts, tool definitions, and accumulated context, so reusing cached segments can meaningfully lower the cumulative cost of a multi-step task. Combined with retrieval-augmented generation, where long retrieved passages are appended to prompts, explicit control over what gets cached can help teams manage the trade-off between content freshness and reuse.

As with most usage-based cloud services, the practical savings will vary with prompt structure, cache hit rates, and how often the cached content changes. Developers evaluating the feature should test against their own traffic patterns rather than assume a fixed reduction. This announcement does not tie the caching capability to a specific new pricing tier beyond the stated goal of lowering inference cost, so the benefit is likely to be realized through reduced token processing on cache hits rather than a separate discount.

Overall, the update is an incremental but practical expansion of Bedrock's foundation-model offering. It widens the range of OpenAI models available on the platform while giving developers a more granular tool for controlling cost. For organizations standardizing on Bedrock as a single point of access to multiple model providers, the addition removes one more reason to maintain separate integrations, and it reinforces AWS's positioning of the service as a provider-neutral hub for model access and optimization.