GPT-5.6 Sol gets a speed option as Bedrock adds controllable caching

OpenAI's July 30 update changes the operating choice around GPT-5.6 rather than describing a new model: Luna's API list price falls 80 percent and Terra's falls 20 percent, while Sol's price remains unchanged. The company also replaces Priority Processing with Fast mode for Sol, saying it can deliver up to 2.5 times the Standard-processing speed at twice the price, with no stated change in intelligence. OpenAI says existing API requests tagged priority will automatically use Fast mode. For teams already using Sol, the immediate decision is therefore whether a latency premium is justified for a particular request path, not whether to requalify a different Sol capability tier.

AWS's technical post adds a separate implementation angle for the GPT-5.6 family on Amazon Bedrock. It describes explicit prompt caching through the OpenAI-compatible Responses API, where an application marks the end of a reusable prefix and supplies a consistent cache key. AWS says cache reads receive a 90 percent discount relative to uncached input, cache writes cost 1.25 times the uncached input rate, and cached prefixes remain reusable for at least 30 minutes. The post identifies repeated system instructions, tool definitions, and reference material in agent loops as the intended use case.

The comparison with the preceding Bedrock setup is more consequential than the headline discount alone. AWS says GPT-5.5 and GPT-5.4 used automatic caching with free cache writes, whereas GPT-5.6 uses a controllable model in which both implicit and explicit writes carry the 1.25-times write rate. AWS estimates that net input cost falls once cache reads account for roughly 20 percent of the tokens flowing through the cache. That makes a migration partly a prompt-layout exercise: a model-ID change alone can preserve compatibility, but it does not establish that a workload's repeated context will produce economical cache reads.

For an agent builder, the practical pattern is to put stable material before an explicit breakpoint, keep variable user input and tool results after it, and reuse the same prompt-cache key across calls that should share a prefix. AWS specifies a minimum 1,024-token cached prefix and allows up to four breakpoints. Its recommended verification is not a generic cost estimate but application logging of `cached_tokens` and `cache_write_tokens` alongside the request ID and cache key. That gives an engineering team a concrete way to distinguish a useful write-once, read-many pattern from a configuration that repeatedly pays cache-write charges.

The limitations are material. OpenAI's announced price reductions apply to Luna and Terra, not Sol, so Fast mode is a latency option rather than a direct Sol price cut. AWS documents Sol in two US East regions and describes cache reuse for at least 30 minutes, which may not fit every deployment or traffic pattern. Neither provider post supplies a task-level comparison of native OpenAI and Bedrock execution for a specific application. Teams should therefore validate quality, latency, cache-read rate, and total token cost on representative production prompts before routing larger shares of traffic.