The Maze: OpenAI has cut GPT-5.6 Luna's price by 80% and Terra's by 20%, only three weeks after launching the model family. Luna now costs $0.20 per million input tokens and $1.20 per million output tokens; Terra costs $2 and $12. The change is bigger than a cheaper API line item. It gives companies more room to assign each step of an AI workflow to the least expensive model that can complete it reliably—while keeping premium reasoning for the decisions where errors cost more than tokens.
High-volume intelligence just became a different budget category. Luna's input and output prices fell from $1 and $6 per million tokens to $0.20 and $1.20. Terra moved from $2.50 and $15 to $2 and $12. Sol remains at $5 and $30. That makes Luna one-tenth of Terra's token price and one-twenty-fifth of Sol's standard price. For commerce teams, the practical candidates are repetitive, well-specified jobs: classifying customer interactions, cleaning product data, enriching catalog attributes, drafting routine variants, checking listings, or handling background steps inside larger agents. A use case that looked too marginal at the old price may now survive a real operating budget.
The operating model is routing, not one-model loyalty. OpenAI's recommended logic is to match intelligence to the outcome, error cost, urgency, and scale. Sol can resolve uncertainty or design the plan. Terra can handle balanced everyday work. Luna can execute well-defined steps, run tests, or process volume. This resembles cloud infrastructure more than a software-seat decision: every task does not need the largest machine. Retailers and platforms that evaluate quality by workflow stage can capture the price cut. Teams that send everything to one default model may simply turn lower unit prices into more undisciplined usage.
The discount came from the serving stack, not a smaller context window. Both Luna and Terra retain a 1.05-million-token context window and broad tool support. OpenAI says the savings came from improvements in model training, request routing, GPU scheduling, production kernels, speculative decoding, cache management, and its agentic harness. GPT-5.6 Sol helped rewrite kernels and run hundreds of draft-model experiments. OpenAI attributes a 20% reduction in end-to-end serving cost to the kernel work and more than a 15% gain in token-generation efficiency to the experiments. Better prompt-prefix caching and less repeated context also matter because an agent may pay the same overhead across dozens of model calls.
Cheap tokens still need expensive discipline. Luna requests above 272,000 input tokens are charged at twice the input rate and 1.5 times the output rate for the entire request. Cache writes carry a premium. Tool calls, retries, evaluation, monitoring, and human review add costs that a token table cannot show. OpenAI's customer examples are encouraging but vendor-selected: Notion says Terra matched GPT-5.5 quality at half the cost per task, while Dust says Luna was 40% faster and 40% cheaper than its previous default on the same agentic work. The real metric is not price per million tokens. It is cost per successful, quality-controlled outcome.
Why it matters: Model economics are moving from procurement into workflow design. Lower Luna and Terra prices make AI more practical in the repetitive layers of ecommerce, but they also raise the value of evaluation, routing, caching, and clear quality thresholds. Operators should map where additional intelligence changes the result and where it merely raises the bill. OpenAI has made the cheap lane much wider. The advantage now goes to teams that know which work belongs in it.
