GPT-5.6 Sol’s latest change is about the harness, not just the model

OpenAI’s July 30 update changes the operating mode around GPT-5.6 Sol. Sol’s standard pricing is unchanged, but Fast mode replaces Priority Processing, is billed at twice the Standard price, and OpenAI says it can deliver up to 2.5 times Standard speed. Existing API requests marked `priority` will route to Fast. The same update cut listed API prices for GPT-5.6 Terra and Luna, rather than Sol, making this chiefly a latency option for the family’s highest tier and a cost reset for its lower tiers.

The immediately preceding research note makes the more consequential technical point: deployment setup can change measured capability. OpenAI compared ARC-AGI-3 results from a generic harness that discarded private reasoning and truncated history with a Responses API harness that retained reasoning and used compaction. It reports a public-set result of 38.3% for Sol with the latter setup versus 13.3% with the official harness, alongside roughly sixfold fewer output tokens. Those are OpenAI’s own measurements under a modified harness, not an independent ranking or evidence of a newly changed base model.

That distinction matters when comparing Sol with GPT-5.5 or GPT-5.4 deployments. AWS’s migration guidance says an existing Bedrock deployment can retain its endpoint, authentication, and Responses API request shape while changing the model ID, but it also calls out two variables that should be tested during the move: reasoning effort and caching. GPT-5.6 adds an `xhigh` effort level, and its Bedrock caching model supports explicit breakpoints with a 30-minute reuse window. A before-and-after evaluation that changes the model, effort level, cache policy, and processing tier at once cannot isolate the source of a gain or cost reduction.

For teams using Sol in long-running tool loops, the practical response is to treat retained reasoning, compaction, cache behavior, and Fast mode as part of the deployment contract. OpenAI recommends the Responses API, retained reasoning, and compaction for workloads seeking the behavior it measured; AWS additionally documents Sol availability in two US regions on Bedrock, while Terra and Luna have a broader regional footprint there. The evidence supports testing these controls against a workload’s own quality, latency, and spend targets. It does not establish that every agent task will see the reported benchmark improvement, or that paying for Fast mode improves answer quality rather than response time.