GPT-5.6 Sol shifts the deployment question from model choice to harness design

OpenAI's July 30 product update made GPT-5.6 Sol's clearest new API change a serving option rather than a new capability claim. Fast mode replaces Priority Processing and, according to OpenAI, can run Sol at up to 2.5 times the Standard speed for twice the price with no claimed change in intelligence. The same update reduced prices for the smaller Luna and Terra tiers while leaving Sol pricing unchanged. That makes the family choice more explicitly a latency-and-routing decision: Sol is the premium path when response time matters enough to justify the surcharge, while lower-cost tiers are positioned for routine, high-volume stages.

A day earlier, OpenAI argued that the surrounding harness can materially change how Sol appears in an agent benchmark. In its ARC-AGI-3 experiment, the company reported a 13.3% score on the public set with the official harness and 38.3% after retaining reasoning and using compaction, alongside six times fewer output tokens. ARC Prize describes ARC-AGI-3 as an interactive benchmark in which agents explore unfamiliar environments without instructions, so it is especially sensitive to how an agent retains observations and plans across a run. The score change is OpenAI's reported result, not an independent ranking of Sol against every competing model or agent stack.

The implementation lesson is concrete. OpenAI says the Responses API can retain reasoning across turns when an application passes the previous response ID, while compaction replaces older context with a summary instead of simply dropping history. Its own recommendation is to use the Responses API, retain reasoning, and use compaction when maximizing performance. Those settings may be most relevant to tool loops, computer-use tasks, and other work where the next action depends on a growing trail of prior actions; a short, stateless request does not automatically inherit the same benefit.

Amazon's July 30 Bedrock documentation adds a separate operational lever for the same family: prompt caching. AWS documents implicit caching by default and explicit cache breakpoints for reusable prefixes, with cache reads billed at a 90% discount, cache writes billed at 1.25 times the uncached input rate, and cached prefixes available for at least 30 minutes. The company says the economics favor write-once, read-many workloads such as agents that repeatedly send system instructions, tools, or reference material. Sol is listed there as the complex-reasoning tier, but its regional availability is narrower than Terra and Luna, which is a practical constraint for teams standardizing on Bedrock.

Teams moving a high-stakes agent should therefore measure at least two interacting levers rather than extrapolate from a headline benchmark: the context policy that retains reasoning and compacts history, and the serving or cache configuration chosen for the workload. OpenAI's account shows why a generic harness can understate Sol in its preferred setup; AWS separately advises replaying representative prompts, logging cache read and write usage, and ramping traffic gradually. Those are provider recommendations, not independent proof of universal gains, but they form an actionable experiment design for deciding whether Sol's latency premium and deployment constraints are justified.