GPT-5.6 Sol on Bedrock makes prompt-cache placement a deployment decision

Amazon Bedrock now supports explicit prompt caching for OpenAI's GPT-5.6 Sol, Terra, and Luna. The feature lets an application mark the end of a reusable prompt prefix, allowing later requests to reuse the preceding content. AWS positions Sol for complex reasoning and agentic coding, so the practical change is aimed most directly at tool-using and multi-turn systems that repeatedly send instructions, tool definitions, or reference material rather than at simple one-shot prompts.

The integration uses the OpenAI Responses API through AWS's bedrock-mantle endpoint. AWS lists Sol in US East (N. Virginia) and US East (Ohio), while Terra and Luna are also listed for US West (Oregon). Existing OpenAI SDK applications can be redirected to the Bedrock endpoint by changing the base URL, model ID, and authentication arrangement, but region availability remains a concrete deployment constraint for teams that need Sol specifically.

The economics favor a stable, long prefix that is written once and read many times. AWS says cached reads receive a 90 percent input-price discount, while cache writes cost 1.25 times the uncached input rate; its guidance says the net input cost falls once reads account for roughly 20 percent of the tokens flowing through the cache. Explicit caching requires a reusable prefix of at least 1,024 tokens, a cache breakpoint at the end of that prefix, and a consistent cache key; cached prefixes remain reusable for at least 30 minutes.

This is a narrower operational improvement than a claim of new model capability. AWS says a GPT-5.5 or GPT-5.4 workload already on Bedrock can move to GPT-5.6 with the same endpoint, authentication pattern, and Responses API shape, mainly changing the model ID. The newer family adds a controllable cache boundary and AWS recommends testing one reasoning-effort level lower during an upgrade because token efficiency can change the quality, latency, and cost tradeoff. OpenAI separately cut Terra and Luna API prices on July 30 while leaving Sol pricing unchanged, a move Axios described as unusually soon after the family's introduction.

The limitation is that cache behavior must be measured rather than assumed. AWS says identical prefixes do not guarantee a hit on every request, and that a real hit requires observing cached_tokens above zero rather than only seeing zero cache-write tokens. Teams should log cached_tokens and cache_write_tokens by request and cache key, because AWS's aggregate CloudWatch metrics do not provide a cache-specific metric. Bedrock also applies per-model, per-region input- and output-token quotas, so caching can reduce input-token pressure but does not remove the need to handle 429 responses with bounded retries.