OpenAI has cut GPT-5.6 Luna and Terra prices and added a faster API route for GPT-5.6 Sol.
The July 30 update says Luna, the fastest and lowest-cost GPT-5.6 tier, now costs 80% less. Terra, the balanced tier for everyday work, now costs 20% less. OpenAI also says the same lower prices affect how usage is counted against paid subscriptions in Codex and ChatGPT Work.
For the high-end Sol tier, OpenAI introduced Fast mode in the API. The company says Fast mode replaces Priority Processing, is backward compatible with requests tagged priority, and can deliver up to 2.5x faster speeds than standard processing at twice the price.
This is a follow-up to the GPT-5.6 launch, where OpenAI positioned Sol, Terra, and Luna as durable capability tiers rather than one model with one price.
The product is routing
The price cut makes GPT-5.6 less like a single upgrade and more like a routing system. Luna can now compete for high-volume tasks where cost decides whether a workflow runs at all. Terra becomes a more natural default for ordinary coding, analysis, and internal assistant work. Sol remains the candidate for hard reasoning, high-stakes agents, or latency-sensitive work when Fast mode is worth paying for.
The refreshed Artificial Analysis data used by The AI Feed reflects the same direction. In today’s model dataset, GPT-5.6 Luna rows list a blended price near $0.174 per million tokens, while GPT-5.6 Terra rows list about $1.74. Sol remains much more expensive at about $4.35 blended, but sits higher in the ranking.
Those are not identical to raw API list prices because the model tracker uses its own blended-price methodology. They still show the buyer problem clearly: model choice now depends on outcome cost, not just answer quality.
Subscription usage is part of the story
OpenAI’s note that Codex and ChatGPT Work usage accounting changes with the new Luna and Terra prices is important. For agent products, list price is only one budget surface. Subscription credits, rate limits, priority lanes, retries, and cache behavior all affect what a team can run.
The practical question is no longer “which GPT-5.6 model is best?” It is “which model should own each class of work?”
