GPT-5.6 Sol turns prompt-cache accounting into an engineering task

The important operational change around GPT-5.6 Sol is not a benchmark table but a new cost and control surface for repeated context. OpenAI lists Sol at $5 per million input tokens and $30 per million output tokens, while Terra is $2.50/$15 and Luna is $1/$6. It also says the family supports explicit prompt-cache breakpoints with a 30-minute minimum cache life. Cache writes are billed at 1.25 times the uncached input rate and cache reads receive a 90% cached-input discount. For systems that resend large instructions, tool schemas, or reference files on each agent step, that turns prompt layout into a cost-sensitive part of application design.

Amazon Bedrock's documentation makes the behavior concrete for its GPT-5.6 endpoint. Sol accepts cache checkpoints after at least 1,024 tokens, permits up to four checkpoints per request, retains them for 30 minutes, and exposes breakpoints on text, image, and file input blocks in the Responses API. AWS describes this as useful when system instructions, tool definitions, and reference files recur while only the newest user input changes. The comparison with GPT-5.5 is consequential: AWS documents automatic caching for the earlier model generation, whereas GPT-5.6 adds explicit breakpoints. Moving an agent unchanged may therefore preserve functionality while changing how engineers should place and measure reusable prefixes.

The cache controls are also not portable across every hosted implementation. Bedrock documents explicit breakpoints for GPT-5.6, but Microsoft's Azure OpenAI documentation says GPT-5.6 caching there remains automatic: developers can provide a prompt-cache key, but Azure does not support `prompt_cache_options` or `prompt_cache_breakpoint`. Microsoft also warns that one changed character in the first 1,024 tokens creates a cache miss and advises keeping stable content first. The practical implication is that a provider-neutral OpenAI-compatible client should not assume a single cache-tuning strategy, even when the model name and high-level API resemble one another.

There are material limits to the apparent savings. AWS says prompt caching is supported only on on-demand inference endpoints, not batch inference, and the minimum checkpoint size means short prompts will not qualify. An open Codex GitHub issue reports a Bedrock workload in which cache writes accounted for 90% of recorded model spend, while explicitly cautioning that its measurements do not establish a universal cause or saving rate. That report is direct but anecdotal evidence, not an independent benchmark. Teams evaluating Sol should therefore instrument cached-token and cache-write usage, test real prefix stability, and compare cold and warm request patterns before treating the discounted-read rate as an achieved saving.