OpenAI Cut Luna's Price 80% and Left Sol Untouched — the GPT-5.6 Range Is Now 25× Wide
GPT-5.6 Luna drops to $0.20 / $1.20 per million tokens from $1 / $6, and Terra to $2 / $12 from $2.50 / $15. Sol is unchanged, and its new Fast option costs twice as much for about 2.5× the speed. OpenAI credits serving efficiency — but a roughly 20% cost improvement explains Terra's cut and comes nowhere near explaining Luna's, which leaves the gap between the cheapest and best model five times wider than it was a month ago.
OpenAI cut the price of GPT-5.6 Luna by 80% today and Terra by 20%. Sol, the flagship, did not move at all — and the new option OpenAI added alongside it costs more.
The cuts are real and immediate. Luna drops to $0.20 per million input tokens and $1.20 per million output, from the $1 / $6 it launched at roughly a month ago. Terra falls to $2 / $12 from $2.50 / $15. OpenAI also says the lower prices are reflected in how usage is counted in Codex and ChatGPT Work, so the same subscription buys more work.
The new price list
| Model | Before (in / out per 1M) | Now |
|---|---|---|
| GPT-5.6 Luna | $1.00 / $6.00 | $0.20 / $1.20 (−80%) |
| GPT-5.6 Terra | $2.50 / $15.00 | $2.00 / $12.00 (−20%) |
| GPT-5.6 Sol | $5.00 / $30.00 | Unchanged |
| Sol — new Fast option | — | ~2.5× speed at 2× the price |
The Fast option replaces what OpenAI had been calling Priority Processing. It is a genuine addition rather than a repackaging of the same speed at a new price, but it points the opposite way from the headline: at the top of the range, OpenAI is selling more performance for more money.
Only one of these cuts is explained by efficiency
OpenAI's stated reason is serving efficiency. The company says gains made during GPT-5.6's own development — including the model rewriting and optimising production code — cut serving costs and improved token generation. Reporting puts those figures at roughly 20% lower serving costs and over 15% better token-generation efficiency; that specific pair of numbers comes from a single secondary account of OpenAI's post rather than a primary document, so treat the precision with some caution.
Take them at face value anyway, and the arithmetic only works for one of the two cuts. A 20% reduction in serving cost maps almost exactly onto Terra's 20% price cut. It comes nowhere near funding Luna's 80%. Cutting a price to a fifth when your costs fell by a fifth means the margin on that tier absorbed the difference.
So these are two different decisions wearing one announcement. Terra's cut is the efficiency story OpenAI is telling. Luna's cut is a competitive one.
The gap between cheapest and best just got five times wider
The more interesting consequence is what happened to the shape of the lineup.
| Distance from Luna to Sol | Before | Now |
|---|---|---|
| Input tokens | 5× ($1 → $5) | 25× ($0.20 → $5) |
| Output tokens | 5× ($6 → $30) | 25× ($1.20 → $30) |
| Cheapest million output tokens | $6.00 | $1.20 |
A month ago the three tiers were evenly spaced, a factor of roughly 2.5 apart at each step. Now the bottom has fallen away from the middle and the top has not moved, so the family spans a 25× price range instead of a 5× one.
That is a deliberate shape. It says the cheap tier is a commodity to be defended on price, and the frontier tier is a differentiated product to be defended on capability. It also means "which GPT-5.6 should I use" is now a much more consequential question than it was yesterday, because getting it wrong costs 25× rather than 5×.
Where the pressure is coming from
Luna is exactly the tier that open-weight models attack. High-volume, latency-sensitive, unglamorous work — summarising, drafting, classification, routing — is where a good-enough model running on rented hardware competes hardest, and where brand loyalty is weakest. Kimi K3's release was the loudest recent example, and it is not the only one.
Enterprise buyers have also become noticeably more reluctant to sign large AI commitments without a clear picture of returns, and cost per task is the easiest lever to demonstrate. Cutting the entry tier to $0.20 per million input tokens is a direct answer to a procurement conversation.
None of this contradicts what we wrote in June about frontier model prices moving up rather than down. It confirms the bifurcation that piece described: the top of the market is getting more expensive while the bottom races toward zero. Today's announcement did both at once, in a single press release.
The counter-case
It would be easy to read an 80% cut as a distress signal, and that would be too quick. Inference costs have genuinely fallen, OpenAI's efficiency claims are plausible on their face, and a company defending the commodity tier of its lineup against open-weight competition is behaving rationally rather than desperately. Losing the cheap tier to open models does not just cost that revenue — it costs the pipeline of developers who start cheap and grow into the expensive tier.
For anyone actually building, this is straightforwardly good news, and the practical takeaway is narrow: re-check your routing. Work that was parked on Terra because Luna felt too weak to be worth the saving is now a 10× cost decision rather than a 2.5× one, and workloads sitting on Sol out of habit deserve a fresh look. It is also worth noting that Sol's unchanged price sits oddly beside today's other Sol story, in which OpenAI argued the model has been under-measured on a reasoning benchmark — a company confident it is winning at the top has little reason to discount there.
You can compare current pricing and capability across vendors on our comparison page and model leaderboard.
Want AI news before everyone else?
The morning's most important AI stories, straight to your inbox. No fluff.