OpenAI has cut the price of GPT-5.6 Luna, its most affordable model, by 80%, just three weeks after the GPT-5.6 family launched. The entry rate drops to 20 cents per million tokens, while Terra falls 20% and the high-end Sol model holds its price. OpenAI credits the cut to efficiency gains on its own infrastructure.
Key Takeaways
- GPT-5.6 Luna drops from 1 dollar to 20 cents per million input tokens, an 80% cut.
- Terra falls 20% to 2 dollars, while the Sol model stays unchanged at 5 dollars.
- OpenAI credits the cut to efficiency gains produced by its own Sol model.
Have an AI Sum Up This Article
ChatGPT80% off the most affordable model
The update sits directly on OpenAI’s official GPT-5.6 page refreshed on July 30. The Luna model, the fastest and cheapest in the lineup, sees its input price fall from 1 dollar to 20 cents per million tokens, and its output from 6 dollars to 1.20.
The cut does not hit the whole lineup the same way. Terra falls 20%, from 2.50 to 2 dollars on input and from 15 to 12 on output. The flagship Sol model holds at 5 dollars input and 30 output. OpenAI keeps the aggressive pricing at the bottom of the range, where price competition bites hardest.
The timing lands as hard as the number. The GPT-5.6 family launched worldwide on July 9, and the cut arrives three weeks later. A price that collapses that fast after a launch says something about the pressure labs face on their margins.
For teams building on the API, the effect is immediate. A task that cost 1 dollar on the best competing models from a year ago now runs at about 6 cents on Luna, with execution OpenAI describes as roughly nine times faster. The capability-to-price ratio flips at once for high-volume use.
The gap it opens is not only against last year’s models. It also widens the distance between OpenAI’s own tiers, nudging developers to think harder about which task truly needs the flagship and which one Luna can carry for a fraction of the cost.
This move is not isolated. It extends a run where OpenAI and Anthropic keep undercutting each other on API pricing. Each cut forces the other side to answer, and the room to maneuver shrinks with every round.
How Sol funded the cut
OpenAI does not frame the cut as a plain commercial move, but as an efficiency gain. The company explains that its own Sol model made its infrastructure more efficient, which frees margin to lower prices without deepening losses.
Two technical levers get the spotlight. A GPU optimization software written by the model itself cuts deployment cost by about 20%. An improvement in token generation, through a speculative decoding technique, adds more than 15% in speed.
There is a self-reinforcing loop in that claim. If each generation of models trims the cost of running the next, OpenAI can keep cutting prices while telling investors the losses are shrinking. Whether the loop holds at scale is the part no pricing page can prove.
The message is as strategic as it is technical. By claiming the AI optimizes its own infrastructure, OpenAI tells a story where price drops become structural rather than one-off. A model that makes the next one cheaper to serve is a promise of a price curve that keeps sliding down.
All three models stay available through ChatGPT Work, the Codex coding tool and the API. A developer can arbitrate finely, pushing heavy workloads toward Luna while keeping Sol for cases that demand the top tier.
More articles on Horizon
- Claude Mythos 5: Anthropic Admits Three Real Breaches
- Claude Opus 5 Cheats to Win a Vending Machine
- Personal AI Agents: Zuckerberg Predicts Billions
Chinese pressure and Microsoft in the backdrop
On the competitive side, the cut does not come from nowhere. OpenAI operates in a market where Chinese providers drag prices down and where serving a model for almost nothing becomes a central selling point.
The threat also comes from inside the Western camp. Microsoft pushes its own MAI models as cheaper alternatives to OpenAI’s offerings, an uncomfortable signal from a longtime partner turned rival on cost.
For a working developer, the near-term read is simple. The cheapest capable model on the market just got cheaper again, and moving a high-volume workload to Luna now carries a smaller penalty than staying on a pricier tier out of habit.
Anthropic plays the same card. The rival leaned on price too, like when Claude Opus 5 came in to match the top of the market at half the cost. The battle shifts from benchmarks to the invoice, and the customer collects the winnings in the short term.
Sustainability is the open question. A price that collapses three weeks after a launch raises doubts about the labs’ ability to profit from models that cost more and more to train. As long as infrastructure efficiency keeps pace, the curve holds. The day it stalls, the price war turns into a war of losses.
Follow the story on Horizon.
