GPT-5.6’s latest change is about operating cost, not a new model tier

OpenAI says that, effective July 30, GPT-5.6 Luna API pricing fell 80% to $0.20 per million input tokens and $1.20 per million output tokens, while Terra fell 20% to $2 and $12 respectively. It also says GPT-5.6 Sol’s new Fast mode replaces Priority Processing, offers up to 2.5 times Standard-processing speed at twice the price, and leaves Sol’s intelligence unchanged. The change is an operational update to the existing GPT-5.6 family rather than evidence of a new release or a change in detector-confirmed release state.

The pricing announcement follows a July 29 engineering account in which OpenAI describes changes across model training, serving and the agentic harness used by Codex and ChatGPT Work. That account says Sol was used within a human-led process to help optimize production kernels and run experiments on a draft model. OpenAI attributes a 20% reduction in end-to-end serving cost to the kernel work and more than 15% higher token-generation efficiency to the resulting speculative-decoding work. Those figures describe OpenAI’s own infrastructure and should be read as provider-reported operational results, not a public benchmark of application quality.

The practical comparison is therefore between tiers and workflow stages, not a universal claim that the cheapest variant wins. OpenAI positions Sol for the most demanding reasoning and coding work, Terra for balanced everyday workloads, and Luna for fast, high-volume work. Its suggested pattern is to use Sol where uncertainty or planning needs the strongest model, then assign well-specified implementation, testing or classification work to lower-cost tiers. The claimed coding comparison with Claude Fable 5 is likewise OpenAI’s own reading of the Artificial Analysis Coding Agent Index and cost estimates, so teams should validate it against their prompts, tools and error tolerance before using it for routing policy.

AWS now makes the cost question more concrete for teams using Amazon Bedrock. Its documentation says all three GPT-5.6 variants are generally available through the OpenAI-compatible Responses API, support a 272K-token context window, and provide implicit or explicit prompt caching. Cached input receives a 90% discount, but cache writes cost 1.25 times the uncached input rate; AWS says the net input-cost benefit begins when cache reads are roughly 20% of tokens flowing through the cache. Explicit caching also requires a stable prefix of at least 1,024 tokens and is retained for at least 30 minutes. That favors agent loops with repeated instructions, tool definitions or reference material, while one-off prompts and constantly changing prefixes may see little benefit.