GPT-5.6 Sol on Bedrock adds explicit prompt caching for repeated agent context

AWS said on July 30 that Amazon Bedrock now offers OpenAI's GPT-5.6 Sol, Terra, and Luna and has added explicit prompt caching through its OpenAI-compatible Responses API. For Sol users, the material change is a serving option rather than a new model capability: callers can set a cache boundary on stable prompt content, reuse that prefix across later requests, and keep the reusable entry available for at least 30 minutes. AWS lists the Sol model ID as openai.gpt-5.6-sol and says Sol is available in US East, N. Virginia and Ohio.

This is a packaging and deployment change, not a substitute for application-level context design. OpenAI's July 30 pricing announcement cut Terra and Luna rates while leaving Sol pricing unchanged, and it says Fast mode for Sol replaces Priority Processing. The Bedrock addition gives Sol deployments a separate lever: reduce repeated-input handling for stable runbooks, tool definitions, and reference material while reserving a frontier tier for work that warrants it. That is an interpretation of the published caching mechanics and product tiering, not a claim that caching improves model quality or benchmark performance.

The comparison with earlier GPT models on Bedrock is concrete. AWS says GPT-5.5 and GPT-5.4 used automatic caching for eligible prefixes of at least 1,024 tokens and did not charge cache writes. GPT-5.6 introduces controllable cache placement, but cache writes are billed at 1.25 times the uncached input rate and cache reads receive a 90 percent input discount. AWS says repeated reads need to make up roughly 20 percent of tokens flowing through the cache before the input-cost calculation becomes favorable. A one-off or highly variable prompt can therefore add configuration and write cost without delivering the intended benefit.

For technical teams, the conservative rollout is to replay a representative prompt set, place explicit breakpoints only after demonstrably stable content, retain a consistent cache key, and inspect cached_tokens alongside cache_write_tokens. AWS recommends that validation before production use and says up to four breakpoints are supported. The limitation is structural: caching avoids recomputing repeated input, not model-output cost, tool execution, or quality regressions from a migration. AWS also says that a move from Chat Completions requires a Responses API migration, while an existing Bedrock GPT-5.5 or GPT-5.4 Responses integration can preserve its endpoint, authentication pattern, and request shape but should test reasoning effort deliberately.