OpenAI has made its smallest GPT-5.6 model dramatically cheaper only three weeks after launch. Starting July 30, GPT-5.6 Luna costs $0.20 per million input tokens and $1.20 per million output tokens—80% below its launch price of $1 and $6.
GPT-5.6 Terra also received a cut, from $2.50 and $15 per million input and output tokens to $2 and $12. Standard GPT-5.6 Sol pricing remains unchanged at $5 and $30.
The new GPT-5.6 price ladder
The change widens the economic distance between OpenAI's three capability tiers. Luna is now one twenty-fifth the standard input price of Sol and one twenty-fifth the output price. Terra remains the middle option for work that needs more capability without paying flagship rates.
That separation matters because production systems rarely need the strongest model for every step. A well-designed workflow might use Sol to resolve ambiguity or make a difficult plan, then hand routine implementation, classification, extraction, testing, or follow-up work to Luna.
At the new rate, processing 100 million Luna input tokens and 20 million output tokens would cost $44 before caching and other service choices. At launch pricing, the same token volume would have cost $220.
Why an 80% cut arrived so quickly
OpenAI attributes the price reduction to efficiency gains across the model, inference stack, routing, production software, and context management. The company says GPT-5.6 Sol helped optimize production kernels and run experiments that reduced serving costs and improved token-generation efficiency.
The commercial timing is just as important. Low-cost open-weight models and aggressive pricing from competing labs are forcing model providers to prove value in deployed workflows, not only on benchmark tables. A model that looks impressive in an evaluation can still be the wrong choice if its token use, latency, retries, and supervision costs make the completed task too expensive.
Fast mode changes the other side of the equation
OpenAI also renamed Priority Processing to Fast mode. For GPT-5.6 Sol, Fast mode promises up to 2.5 times Standard processing speed at twice the token price. Existing requests tagged with the priority service tier remain compatible, so the change is not intended to break current integrations.
This creates a clearer operational choice: pay less for high-volume work with Luna, use Terra when the quality balance warrants it, keep Sol for the hardest decisions, and pay a premium for Fast mode only when latency has measurable business value.
What this means for small product teams
For independent studios such as SunMarc App Labs, an 80% reduction can move features out of the “interesting but uneconomical” column. Background agents, support triage, document processing, content classification, in-app assistance, and multi-step automations can run more often without forcing a subscription price that users will reject.
It also rewards better system design. The winning architecture may not send every request to one default model. Teams can route by task difficulty, set budgets, exploit prompt caching, measure accepted outcomes, and escalate only the cases where a more expensive model changes the result.
Lower token prices do not remove the need for evaluations. If a cheaper model needs more retries, produces longer answers, or creates more review work, its real cost advantage can shrink. The useful metric is still cost per successful task—not the price printed beside a model name.
The larger signal
This unusually fast price cut shows that frontier AI competition is becoming an efficiency contest. Capability still matters, but providers are increasingly competing on how much dependable work customers can finish for each dollar and each second of latency.
That shift expands the addressable market for AI features. When capable tool use becomes inexpensive enough to disappear into ordinary product economics, AI stops being a premium showcase and starts behaving like a standard software utility.