GPT-5.6 Sol on Bedrock makes cache discipline a first-class agent cost control

AWS documents GPT-5.6 Sol, Terra, and Luna through its OpenAI-compatible Responses API on the Bedrock Mantle endpoint. For Sol, AWS positions the service around autonomous coding, security research, scientific analysis, and deep multi-step reasoning. The notable new operational feature is explicit prompt caching: developers can mark the reusable end of a prompt prefix rather than relying only on an automatically placed cache boundary. That makes prompt layout an application-level decision for teams that repeatedly send policy text, tool schemas, repositories, or reference material to an agent.

On Bedrock, cache reads for GPT-5.6 are billed at a 90 percent discount from uncached input, while cache writes cost 1.25 times the uncached input rate. AWS says a workload reaches lower net input cost when cache reads account for roughly 20 percent of the tokens flowing through the cache. That is most relevant to write-once, read-many agent loops, not to a single isolated completion. The platform exposes `cached_tokens` and `cache_write_tokens` on every response, so teams can verify whether a supposedly stable prefix is actually being reused instead of treating caching as an unmeasured optimization.

This is a different lever from OpenAI's July 30 price changes. OpenAI lowered GPT-5.6 Terra and Luna API prices, left Sol pricing unchanged, and replaced Priority Processing for Sol with Fast mode at twice the Standard price. Bedrock caching therefore gives Sol users a way to reduce the repeated-context portion of a workflow without claiming a lower base Sol rate. Inference from the published pricing and cache mechanics is straightforward: a workflow can reserve Sol for uncertain planning or difficult debugging, then reduce its context overhead by keeping the expensive shared prefix stable and routing routine stages to cheaper family members when its own evaluations support that split.

The constraints matter. A cacheable prefix must contain at least 1,024 tokens, cached content is available for at least 30 minutes, and a request can set up to four explicit breakpoints. Sol is listed only in AWS US East regions, whereas Terra and Luna also appear in US West (Oregon). Short prompts, frequently rewritten instructions, and work spread beyond the cache window will not obtain the same benefit. Teams should begin with implicit caching, instrument the two cache usage fields, then add an explicit breakpoint only after confirming a repeatable prefix and a favorable write-once, read-many pattern in production-like traffic.