OpenAI cuts GPT-5.6 Luna 80% — reprice your cheap tier

Three weeks after shipping GPT-5.6, OpenAI cut the price on two of the three tiers. The GPT-5.6 Luna price cut is the headline: 80% off, from $1 per million input tokens and $6 per million output to $0.20 and $1.20. Terra drops 20%, from $2.50/$15 to $2/$12. Sol, the top tier, is unchanged. If you sized an AI workload against the July 9 launch prices, that spreadsheet lasted less than a month.

What actually happened

Per CNBC, OpenAI attributed the cuts to efficiency gains found while building GPT-5.6 — the company says it reduced the end-to-end cost of serving the model by 20% and improved token-generation efficiency by more than 15%. The framing OpenAI is going with is that the model helped optimize its own serving stack.

The second change is easy to miss and will cost somebody a production incident. OpenAI is introducing a Fast mode in the API for GPT-5.6 Sol — up to 2.5x faster than standard at twice the price — and it replaces Priority Processing. If you're paying for Priority Processing today, that's a SKU swap, not a new option sitting next to your old one.

Yahoo Finance reads the cuts as a response to enterprise cost sensitivity. Sure. The operator reading is simpler: token prices are a dial the vendor turns, and they turned it 80% in twenty-one days.

Why a cheap-tier price cut matters for your business

An 80% cut doesn't save you 80%. It moves the line between what you route to the cheap model and what you send to the expensive one — and that line is an architecture decision, not a billing one.

Concretely: the classification step, the extraction pass, the first-draft summarization, the "is this email worth a human" triage. Most teams put that work on a mid-tier model because the quality gap felt worth the money. At $0.20 input, the math changes — you can now afford to run the cheap model twice and have a second pass check the first, which usually beats one call to a better model at the same total cost.

What you should not do is chase the number. Re-run cost per completed task, including retries, not cost per million tokens. And build the routing layer so this is a config change: if a vendor can cut 80% in three weeks, they can raise it, deprecate the tier, or replace your latency SKU out from under you. Model choice belongs in configuration, not scattered across your codebase.

Sources: CNBC, Yahoo Finance.