GPT-5.6 Sol on Bedrock gets explicit caching, but the savings require real reuse

AWS has published an implementation guide for explicit prompt caching with OpenAI's GPT-5.6 family on Amazon Bedrock. GPT-5.6 Sol is available through Bedrock's OpenAI-compatible Responses API at the bedrock-mantle endpoint, alongside the Terra and Luna tiers. The material change for technical users is not a new benchmark claim: it is the ability to place a cache boundary around a reusable prompt prefix, such as stable instructions, tool definitions, or reference material, while later request content changes.

The timing matters because AWS first posted the GPT-5.6 Bedrock availability notice on July 13, then followed with practical migration and caching guidance on July 24 and July 30. On Bedrock, Sol is positioned for deep multi-step reasoning, autonomous coding, security research, and scientific analysis; Terra is the balanced tier, while Luna targets high-volume, latency-sensitive work. The integration preserves the Responses API shape, but Sol is currently limited to the two US East regions AWS lists, unlike Terra and Luna, which are also available in US West (Oregon).

This is a meaningful difference from the GPT-5.4 and GPT-5.5 cache behavior AWS describes. Those generations automatically cached eligible prefixes of at least 1,024 tokens, with no additional cache-write charge. GPT-5.6 keeps implicit caching as the default but adds an explicit mode in which the application chooses the boundary. AWS charges cache writes at 1.25 times the ordinary input-token rate and cache reads at a 90 percent discount, so the lower effective input cost is an outcome of a write-once, read-many workload rather than a property of every request.

For an agent team, the practical test is straightforward: put only stable material before the breakpoint, use a consistent cache key, and inspect cached_tokens and cache_write_tokens in production. A reusable prefix must be at least 1,024 tokens, there can be up to four breakpoints, and the documented minimum cache lifetime is 30 minutes. Short prompts, frequently changing prefixes, or workloads that write caches without enough later reads may not benefit and can incur the write premium. AWS's own guidance therefore supports a gradual rollout with representative prompts and cache telemetry, rather than treating the advertised discount as a guaranteed application-level saving.