#OpenAI #GPT56 #Inference #AgentEconomics
OpenAI finally made visible what usually stays inside the data center: optimization of the model, inference, and agent harness no longer only lowers internal cost. It changes the public economics of the product.
In GPT-5.6, Luna API pricing dropped by 80%, Terra became 20% cheaper, and Sol gained Fast mode: the same intelligence level in an accelerated mode at a higher price. In Codex and ChatGPT Work, this shows up not as a discount banner, but as lower credit usage for Terra and Luna.
The architectural conclusion matters more than the numbers. The model family starts to behave like a resource planner: Sol is for uncertainty and expensive branches, Terra covers normal daily work, and Luna pulls through high-volume, well-specified steps. If a workflow has evals, you do not have to “move it to the new model”; you can route it by cost of error, urgency, and scale.
For New Runtime, this maps directly onto the next version of the agent runtime: classify the work first, then choose the model, not the other way around. Savings become a property of the execution graph.