GPT-5.6 Sol on Bedrock Makes Prompt Caching a Deliberate Agent Design Choice

Amazon Bedrock’s GPT-5.6 documentation identifies GPT-5.6 Sol as supporting explicit prompt-cache breakpoints through the Responses API. A cacheable checkpoint needs at least 1,024 tokens; Sol permits as many as four checkpoints per request and lists a 30-minute TTL. That gives an agent builder a defined place to put persistent system policy, tool definitions, or reference material before a breakpoint while leaving the user request and newly returned tool results outside it.

AWS describes the accounting as write-once, read-many: cache reads receive a 90 percent discount relative to uncached input, while cache writes cost 1.25 times the normal input rate. Its guidance estimates that the input-cost benefit begins once cache reads account for roughly 20 percent of the tokens passing through the cache. This is most relevant to long-running support, retrieval, and tool-calling flows where a large prompt prefix survives across several requests, not to isolated short prompts.

The design differs from the GPT-5.5 path AWS describes. Earlier GPT caching on Bedrock was automatic for eligible prefixes of at least 1,024 tokens and its cache writes were free; GPT-5.6 adds explicit placement and the cache-write charge. For a stable prefix followed by a changing suffix, that trade is useful because a fixed prompt_cache_key and a breakpoint after the static material make the reusable portion visible and controllable instead of relying on an automatically chosen boundary.

The integration path also has practical constraints. AWS serves the GPT-5.6 family through an OpenAI-compatible Responses API on its bedrock-mantle endpoint, so an existing Responses client can retain its broad request shape while replacing the endpoint, authentication, and model ID. A Chat Completions client needs a more substantive port to Responses input, output, and tool-call formats. Sol is listed only in the two US East regions in AWS’s launch guidance, whereas Terra and Luna are also listed for US West, so regional deployment requirements can determine whether Sol is an option at all.

The documentation does not turn a billing mechanism into a universal performance result. AWS says applications should inspect cached_tokens and cache_write_tokens on every response, and its migration guide explicitly recommends testing reasoning effort during an upgrade. OpenAI’s separate ARC-AGI-3 write-up likewise attributes its reported Sol improvement to retained reasoning and compaction in the evaluation harness. The defensible operational conclusion is to benchmark the complete agent loop, including cache-hit behavior and model settings, rather than treating a cache discount or a benchmark configuration as a promised workload outcome.