GPT-5.6 Sol on Bedrock turns prompt caching into an engineering choice

Amazon Bedrock's current documentation describes a concrete workflow change for GPT-5.6 Sol: prompt caching can use explicit breakpoints in Responses API requests rather than relying only on automatic prefix reuse. The documentation lists Sol's model ID, a 1,024-token minimum cacheable prefix, up to four checkpoints per request, and a 30-minute minimum cache life. For agent loops that repeatedly send long system instructions, tool definitions, or reference files, that gives the application a deliberate place to end a reusable prefix instead of treating cache behavior as wholly invisible infrastructure.

The comparison with the preceding generation is operational, not merely cosmetic. AWS documents GPT-5.5 and earlier OpenAI models on the same endpoint as using automatic caching for eligible prefixes, with no special parameters and no cache-write fee. For GPT-5.6, AWS documents implicit mode as the default while also allowing explicit breakpoints, so migration does not require every application to add markers immediately. The tradeoff is that explicit mode creates a new prompt-design responsibility: stable material must remain ahead of the breakpoint, while changing user input belongs after it if a later request is to reuse the prefix.

The cost case should be tested rather than assumed from the cache-read discount. OpenAI states that GPT-5.6 cache writes are billed at 1.25 times the uncached input rate and cache reads receive a 90 percent cached-input discount. AWS recommends tracking both `cache_write_tokens` and `cached_tokens` from response usage, and says its endpoint does not provide a cache-specific CloudWatch metric. A workload with a short-lived or frequently changing prefix can therefore incur writes without enough subsequent reads to offset them; a repeated agent loop with a large stable prefix has a clearer path to benefit.

For teams using the OpenAI SDK through AWS, the implementation is also specific to Bedrock's `bedrock-mantle` endpoint and Responses API. AWS says access requires an account with inference permission for that endpoint, and its example material treats response usage as the authoritative place to measure cache activity. The documentation establishes the available controls and constraints, not a guaranteed latency or cost outcome for any application. A practical rollout is to keep the existing prompt structure as a baseline, add one explicit breakpoint around a genuinely stable prefix, then compare cache reads, writes, token spend, and task quality on representative traffic before expanding the pattern.