GPT-5.6 Sol's latest gains are as much about the agent harness as the model

OpenAI's July 29 and July 30 updates make the GPT-5.6 Sol story less about a single headline benchmark and more about the engineering around a model. In an ARC-AGI-3 post, OpenAI says that retaining private reasoning across turns and using compaction raised Sol's public-set RHAE from 13.3% to 38.3%, while cutting output tokens sixfold. The following day, it said it would lower the prices of GPT-5.6 Terra and Luna and add a faster Sol API mode. These are separate changes, but together they put the full agent loop, including state, context handling, latency, and tier selection, at the center of the practical discussion.

The ARC result matters because OpenAI says the initial harness discarded private reasoning after each action and used a rolling truncation window. Its revised setup used the Responses API to retain reasoning through the prior response and replaced truncation with compaction. OpenAI's developer documentation describes compaction as carrying needed prior state and reasoning into subsequent runs with fewer tokens. The useful technical conclusion is not that Sol became intrinsically three times stronger, but that long-running-agent comparisons need to disclose their state-management choices.

For an engineering team, the resulting design pattern is explicit routing rather than treating every request as a Sol request. OpenAI's own example is to use Sol to resolve uncertainty and define a plan, then use Luna for well-specified implementation, tests, and evaluation. As of July 30, OpenAI lists Terra at $2 input and $12 output per million API tokens, Luna at $0.20 input and $1.20 output per million, and unchanged Sol pricing; Sol's Fast mode is advertised at up to 2.5 times Standard speed for twice the price. Whether that routing preserves product quality remains a workload-specific question that teams should measure with their own evals.

The limitations are substantial. The ARC-AGI-3 score change is OpenAI's comparison of two harness configurations on a public task set, not an independent cross-model verdict, and the provider itself notes that benchmarks also reflect API settings, harness design, and prompting. The pricing and serving-efficiency claims are likewise provider claims, although Axios independently reported the Terra and Luna reductions and that Sol did not receive a price cut. For users, the defensible takeaway is narrower: preserve useful agent state, test compaction in long workflows, and compare quality, latency, and cost at the system level before attributing an agent result to the base model alone.