GPT-5.6 Sol’s New Cost-and-Context Playbook

OpenAI’s July 30 update changes the operating economics around GPT-5.6 without changing Sol’s listed token price. The company cut Terra to $2 input and $12 output per million API tokens, and Luna to $0.20 and $1.20, while saying their usage will consume fewer subscription credits. For Sol, OpenAI replaced Priority Processing with Fast mode. It says Fast mode costs twice Standard processing, can provide up to 2.5 times the speed, does not change intelligence, and accepts existing priority-tagged API requests. That is a configuration change with a clear latency premium, not evidence that every Sol workload will become cheaper on its own.

OpenAI’s related ARC-AGI-3 analysis makes the harness part of the performance story. In the company’s reported public-set experiment, GPT-5.6 Sol moved from 13.3% with the official harness to 38.3% when retained reasoning and compaction were enabled, while using six times fewer output tokens. OpenAI attributes the difference to a baseline harness that discarded private reasoning after actions and dropped older history through rolling truncation. The result is useful operational evidence, but it remains a vendor-reported, harness-specific comparison rather than an independently replicated measure of Sol’s general capability.

AWS supplies a concrete deployment example for the same family. Its Bedrock documentation places Sol, Terra, and Luna behind an OpenAI-compatible Responses API and describes explicit prompt caching for repeated system instructions, tool definitions, and reference material. Sol is documented for the two US East regions, while Terra and Luna have an additional Oregon region. Cache reads receive a 90% discount relative to uncached input, but cache writes cost 1.25 times the uncached input rate; the design therefore favors repeated long prefixes over isolated calls. A reusable prefix must contain at least 1,024 tokens, and cached content is retained for at least 30 minutes.

For existing Bedrock users, GPT-5.6 is not entirely a model-ID substitution. AWS says the endpoint, authentication, and Responses API request shape remain the same for GPT-5.5 or GPT-5.4 migrations, but the cache contract changes. Earlier eligible prefixes used automatic caching with free cache writes; GPT-5.6 adds controllable implicit or explicit modes and the 1.25-times write charge. An explicit breakpoint can preserve a static prefix while later user input changes, but the application must keep a consistent cache key and should monitor both cached tokens and cache-write tokens. Cost forecasts that ignore the cold-write phase can therefore misstate the result.

Technical teams should test Sol with the production context strategy they intend to run, not only with a generic benchmark harness. OpenAI’s own workflow example assigns Sol to resolving uncertainty and planning, then assigns well-specified implementation and test work to Luna; that is useful vendor guidance, not a universal routing rule. Compare retained-reasoning and compaction settings under a fixed quality target, record cache-read and cache-write behavior for repeated prompts, and account for region limits and the 30-minute cache window. Fast mode should likewise be evaluated against its stated two-times price rather than treated as a free speedup. Those measurements separate model claims from the behavior of a particular agent stack.