GPT-5.6 Sol’s practical update is in the harness, not just the model

Amazon’s July 30 Bedrock update adds explicit prompt caching controls for OpenAI’s GPT-5.6 family, including GPT-5.6 Sol. Developers can mark the end of a reusable prompt prefix, give that prefix a stable cache key, and keep changing content after the boundary. This is a meaningful operational change for agents with long system prompts, retrieval context, or tool instructions: it makes the cache boundary a design choice rather than an opaque platform behavior. It does not, by itself, establish anything about the model’s release state or its quality on a task.

The timing matters because OpenAI has also argued that the surrounding harness materially changes what GPT-5.6 Sol can do. In its ARC-AGI-3 experiment, OpenAI reported that retaining reasoning and using compaction raised Sol’s public-set score from 13.3% to 38.3% while reducing output tokens sixfold. ARC Prize’s independently inspectable results page, however, lists Sol Max at 7.78% on its verified ARC-AGI-3 score, with Terra at 0.80% and Luna at 0.18%. The figures are not a contradiction: they describe different setups. The useful lesson for technical users is that model comparisons which discard hidden reasoning or truncate history can measure a harness limitation as much as a model limitation.

The Bedrock change makes that lesson concrete, but it also changes the cost model relative to GPT-5.4 and GPT-5.5. AWS says GPT-5.6 cache reads receive a 90% discount, while cache writes cost 1.25 times the ordinary input-token rate; it estimates a net input-cost benefit once cache reads account for roughly 20% of tokens flowing through the cache. Earlier Bedrock versions used automatic caching with free writes. GPT-5.6 supports up to four explicit breakpoints, requires each cached prefix to contain at least 1,024 tokens, and keeps explicit prefixes available for at least 30 minutes. That favors write-once, read-many workloads over prompts that change throughout.

For an existing Bedrock deployment, the lowest-risk migration is not simply a model-ID swap. AWS recommends placing an explicit breakpoint after genuinely static content and monitoring cache-read and cache-write fields in production; OpenAI separately recommends retaining reasoning and using compaction through the Responses API for long-running work. Teams should test reasoning effort one level lower as AWS suggests, then compare task success, latency, cache reuse, and total cost rather than treating a benchmark number as a deployment forecast. The price cuts announced July 30 apply to Terra and Luna, not Sol, so Sol’s economic case remains strongest where its extra reasoning and faster processing are demonstrably worth the premium.