GPT-5.6 pricing shifts the routing calculus, while Sol becomes the latency tier

OpenAI's July 30 GPT-5.6 update changes the economics around Sol more than it changes Sol itself. The company cut API pricing for Terra to $2 per million input tokens and $12 per million output tokens, and for Luna to $0.20 and $1.20 respectively; it said Sol pricing remains unchanged. At the same time, OpenAI replaced Priority Processing with Fast mode for Sol, promising up to 2.5 times Standard-processing speed at twice the price and retaining compatibility for requests previously tagged as priority. The immediate product change is therefore a wider price and latency spread inside the GPT-5.6 family, not a new claim of additional Sol capability.

OpenAI's July 29 ARC-AGI-3 write-up provides useful context for why an API configuration can matter as much as the model label in agent workloads. The company reported that GPT-5.6 Sol scored 13.3% on the benchmark's public set in the official harness and 38.3% after it retained reasoning across turns and used compaction, while using fewer output tokens. OpenAI attributes the difference to the original harness discarding private reasoning after each action and dropping older context as a rolling window filled. Those are vendor-reported results rather than an independent benchmark finding, but they make the operational point concrete: a model that must repeatedly reconstruct its plan can be slower and more expensive even when the underlying weights have not changed.

The practical comparison is not simply Sol versus a predecessor. OpenAI says Terra performs as well as GPT-5.5 on its intelligence benchmarks at half the price, while Luna is positioned as the fastest and least expensive member of the family. Its pricing post proposes a workflow split in which Sol handles uncertainty and planning, then Luna handles specified implementation, testing, and evaluation. Sol's new option is correspondingly a latency purchase, not a discount: Fast mode costs twice Standard processing and is intended for cases where response time justifies that premium. That makes explicit routing and task-level evaluation more important than selecting one default model for an entire agent pipeline.

For technical teams, the sensible next step is to measure a representative workflow across the three tiers, including retries, tool calls, context growth, and the price of latency, rather than extrapolating from token rates alone. OpenAI recommends its Responses API, retained reasoning, and compaction for developers trying to maximize performance, and its ARC analysis shows why those settings should be held constant in comparisons. What remains uncertain is how broadly the reported gains transfer outside OpenAI's chosen benchmark and harness, and whether the lower-cost models preserve task quality on a particular production workload. The announced price cuts and Fast mode are verifiable product changes; the performance comparisons and customer examples remain OpenAI's own evidence and should be tested independently before they drive a migration.