OpenAI Cut GPT-5.6 Luna by 80% and Left Sol at $30/M Output

OpenAI cut the floor of its GPT-5.6 API by 80% three weeks after the model family reached general availability, and left the top tier untouched. A 2x-priced Fast mode for that top tier is now layered on top. The price curve across the family widened; it did not compress.

The cuts took effect on July 30, 2026, per OpenAI's "Advancing the price-performance frontier with GPT-5.6" post. GPT-5.6 reached general availability on July 9, 2026. The family ships in three developer-facing tiers on OpenAI's API (the developer pricing surface): Luna, the cheapest and fastest; Terra, the mid tier; and Sol, the top tier. The "tokens" the API charges against are chunks of text the model reads (input) or writes (output), priced per million.

Luna, the cheapest tier, dropped to $0.20 per million input tokens from $1.00, and to $1.20 per million output tokens from $6.00. That is an 80% cut on both legs, per OpenAI. Terra, the mid tier, fell to $2.00 per million input from $2.50, and to $12 per million output from $15, a 20% cut on both legs. Sol, the top tier, is unchanged at $5 per million input and $30 per million output, per the GPT-5.6 product page.

Read those four moves together. The floor fell by 80%. The middle fell by 20%. The ceiling did not move. The price curve across GPT-5.6 widened.

OpenAI attributes the lower prices to "improvements across models, inference systems, hardware usage, production software, and context management," per the same post. The company adds that GPT-5.6 Sol optimizations cut its own serving costs by 20% and improved token-generation efficiency by more than 15%. Those are OpenAI's self-reported numbers, not independently verified.

A team running Luna at the old $1.00 per million input and $6.00 per million output now pays the same workload for roughly a fifth of the bill. A team routing hard problems to Sol at $5 per million input and $30 per million output pays the same as it did on day one. The gap between a Luna call and a Sol call, per million output tokens, is now $1.20 against $30, a 25x spread, up from 5x on launch day. That is the unit-economics change under the 80% headline.

OpenAI replaced the API's old "Priority Processing" option with a "Fast mode" for Sol that runs up to 2.5x faster than standard processing and costs 2x the standard Sol price. The model's intelligence does not change; the speed and the bill do. A Fast-mode Sol call now runs at $10 per million input and $60 per million output. Fast mode is the only price line that moved up.

What the cut does not touch is also part of the story. ChatGPT and Codex subscription prices are unchanged, and the credit budgets in ChatGPT Work and Codex are unchanged. Terra and Luna now consume fewer credits in those products because their token prices fell, but the consumer sticker did not change this month. The developer API is the surface that moved; the consumer products did not.

This is not a "price war": the cheapest tier fell while the top tier did not, the opposite of a race to the bottom across the lineup. It is also not a DeepSeek-shock reprice. OpenAI credits its own inference and software work, and the cuts are concentrated in the entry tier it just launched. The honest read is narrower: OpenAI is repricing the floor of a brand-new model family three weeks in, holding the ceiling, and adding a 2x-priced Fast lane on top.

The next repricing is the data point. If Sol's list price is still $5/$30 a quarter from now, this is a builder-friendly widening rather than a generalized reprice. If Sol falls too, this is the start of something broader.