GPT-5.6 Sol on Bedrock gains controllable prompt caching

Amazon Bedrock's July 30 technical guide documents explicit prompt caching for OpenAI GPT-5.6 Sol, Terra, and Luna through its OpenAI-compatible Responses API. Rather than leaving cache placement entirely to the platform, a developer can mark the end of a reusable prompt prefix and supply a stable cache key. AWS says cache reads receive a 90 percent discount versus uncached input and remain reusable for at least 30 minutes. The practical target is an agent or application that repeatedly sends the same system instructions, tool definitions, or reference material while only the user request and accumulated tool results change.

The change is an operational addition to the GPT-5.6 family on Bedrock, not a new benchmark result. AWS's July 24 integration guide described Sol as the tier for autonomous coding, security research, scientific analysis, and deep multi-step reasoning, with the three models accessed through a Bedrock endpoint that works with the OpenAI SDK. OpenAI then announced lower API rates for Terra and Luna on July 30 while saying Sol pricing was unchanged. That sequence matters because caching can alter the effective input cost of a workload even when its selected model's listed token price does not change.

For teams already using earlier GPT models on Bedrock, AWS describes the basic GPT-5.5 or GPT-5.4 migration as a model-ID update while retaining the endpoint, authentication, and Responses API shape. The caching feature adds a more consequential design choice: implicit mode lets Bedrock select cache breakpoints, while explicit mode lets the application place them. That makes Sol's economics depend less on the headline model price and more on whether a deployment separates stable instructions from changing conversation, retrieved material, and tool outputs.

The limitation is that this is not a universal discount. AWS requires a reusable prefix of at least 1,024 tokens, bills cache writes at 1.25 times the uncached input rate, and permits up to four explicit breakpoints; its own guidance estimates that cache reads need to represent roughly 20 percent of tokens flowing through the cache before net input cost falls. Sol is also listed only in US East (N. Virginia) and US East (Ohio) in the cited Bedrock guide. Technical teams should log cached and written token counts before treating the feature as a savings assumption, and this source set does not establish an independent quality or latency advantage.