The gpt-5.6 line runs from $0.20 per million input tokens on luna to $12.50 on cyber, with output at $1.20 and $75.00 respectively, read from the OpenAI pricing page on 5 September 2026.
One line of that grid carries a date. OpenAI marks the sol rates as promotional and says they hold until 30 November 2026, when the promotion ends, so the $4.00 and $20.00 below are a temporary price with a fixed end date rather than a settled one. Budget sol at what it costs after that date, and treat the difference as money you are saving until then.
Four models share the 5.6 name and the top one costs sixty-two times the bottom one on output, which makes choosing the right tier inside the family a bigger decision than choosing OpenAI over anyone else. Output is six times input on luna, terra and cyber, and five times on sol; cached input is a tenth on all four.
What the gpt-5.6 line costs, tier by tier
One family, four prices, and the spread between them is wider than the spread between most providers.
| Model | Input | Cached input | Output |
|---|---|---|---|
| gpt-5.6-luna | $0.20 | $0.02 | $1.20 |
| gpt-5.6-terra | $2.00 | $0.20 | $12.00 |
| gpt-5.6-sol | $4.00 | $0.40 | $20.00 |
| gpt-5.6-cyber | $12.50 | $1.25 | $75.00 |
The gpt-5.6 line, from $0.20 to $12.50 on input, read 5 September 2026
A cached read is a tenth of a fresh one on all four models, and that part is perfectly regular. The output multiplier is not: it is six times input on luna, terra and cyber, but five times on sol, which OpenAI cut to $4.00 and $20.00 after this page first went up.
That one exception matters more than it looks. Moving from luna to terra multiplies every column by ten, and from luna to cyber by sixty-two and a half, so those two comparisons hold whatever your token mix. sol is twenty times luna on input but only sixteen and two thirds on output, so where it lands against the others depends on how much you generate. The regularity also shifts on long prompts, and the section below explains where.
Sixty-two times, between two models sold under the same name. Picking the wrong tier inside gpt-5.6 costs more than picking the wrong provider.
The output multiplier, and the tier that breaks it
At six times input on three of the four tiers, the output multiplier is the highest among the providers we track. Anthropic charges five times more for output across the Claude line, xAI three times on grok-4.6, and Mistral three times on its flagship. sol, at five times, is the one tier that lands on Anthropic’s ratio rather than OpenAI’s.
That matters most on generative work. A summariser that reads a lot and writes a little is well served here; an agent that produces long plans, code or drafts pays the six-times line on most of its bill. Before comparing headline input rates across providers, work out your own output-to-input split, because it can reverse the ranking entirely.
The cache line, identical across the family
On a workload that resends the same system prompt or document on every call, that tenth is worth more than a tier change on the input side. It does nothing for output, which on this line is where most of the money goes, so caching alone will not rescue a bill dominated by long generations.
Unlike Anthropic, OpenAI does not publish separate cache write lines on this page, so the arithmetic is simpler: fresh input, cached input, output. We work through what the extra Anthropic columns change on our Claude API pricing page, cache included.
Long prompts are billed on a second grid
Three of the four tiers carry a separate long-context rate, and it does not scale the way the standard grid does.
| Model | Standard, in / out | Long context, in / out |
|---|---|---|
| gpt-5.6-luna | $0.20 / $1.20 | $0.40 / $1.80 |
| gpt-5.6-terra | $2.00 / $12.00 | $4.00 / $18.00 |
| gpt-5.6-sol | $4.00 / $20.00 | $8.00 / $30.00 |
| gpt-5.6-cyber | $12.50 / $75.00 | no long context rate published |
Input doubles and output rises by half on each of the three, so every output multiplier drops: six times becomes four and a half on luna and terra, and five times becomes three and three quarters on sol. That is the one place where the family’s arithmetic is not uniform, and it works in your favour on generative work: the longer the prompt, the less punishing the output line becomes relative to input.
Two things the page does not tell you. It does not publish the token threshold at which the long grid takes over, and gpt-5.6-cyber has no long-context row at all, only dashes. gpt-5.2 has none either. So the most expensive model in the family and the previous generation are both billed on a single grid whatever you send them.
This is the same structural idea as the 200,000-token threshold on grok-4.6, except that xAI doubles every line while OpenAI doubles input and adds only half to output.
What a month actually costs
Ten million input and two million output tokens, priced on each tier with ninety percent of input served from cache.
| Model | A month at that volume |
|---|---|
| gpt-5.6-luna | $2.78 |
| gpt-5.6-terra | $27.80 |
| gpt-5.6-sol | $47.60 |
| gpt-5.6-cyber | $173.75 |
terra is exactly ten times luna and cyber exactly sixty-two and a half times, on every column, so those two ratios hold at any volume. sol is the exception since its cut: twenty times luna on input, sixteen and two thirds on output, which lands it near seventeen times on the mix above and moves with yours. Our LLM cost calculator runs the same volume across every provider we track.
The rest of the GPT catalogue is still there
The 5.6 family is no longer OpenAI’s newest line: on 6 October 2026 the pricing page opens on gpt-6-astra, gpt-6.1-sol and gpt-6-luna, and gpt-6-luna at $0.10 and $0.50 now undercuts luna, while terra and luna are priced on their own model pages. Several older models remain available and cheap. gpt-5-nano sits at $0.05 and $0.40, below anything in the 5.6 family, and gpt-5-mini is at $0.25 and $2.00, both now marked deprecated by OpenAI. We cover both on our gpt-5-mini and nano page. The gpt-5.2 generation is still listed at $1.75 and $14.00, with its own page here. For those seeking more advanced capabilities, compare these to the GPT-6 flagship above them, or explore the GPT-6 Sol and Luna grid for additional options.
None of those are marked deprecated on the pricing page. If a cheaper older model does your task, the grid gives you no reason to move.
How you buy it changes the price
Everything above is the first-party rate. Azure runs its own grid on the same models, adding ten percent for a US or EU data zone and twenty percent for APAC or India, which we measured on our Azure OpenAI page. Routers add their own margin on top of the provider grid, which we measured on OpenRouter.
On a line where output costs six times input, a few percent of markup lands hardest on exactly the traffic you generate most of. Check the router’s output rate specifically, not its headline.
Frequently asked questions
How much does gpt-5.6 cost per million tokens?
It depends on the tier: luna is $0.20 input and $1.20 output, terra $2.00 and $12.00, sol $4.00 and $20.00, cyber $12.50 and $75.00. Cached input is a tenth of fresh input on all four. Read from the OpenAI pricing page on 5 September 2026, with the sol end date re-read on 10 November 2026. The sol rates are promotional and OpenAI offers them until 30 November 2026, when the promotion ends.
Which gpt-5.6 model is the cheapest?
luna, at $0.20 input and $1.20 output. Outside the 5.6 family, gpt-5-nano is cheaper still at $0.05 and $0.40.
What is gpt-5.6-cyber and why is it so expensive?
It is the top of the 5.6 line, at $12.50 input and $75.00 output. The pricing page does not explain what separates it beyond listing it in its own section, so we price it and leave the capability question alone.
Why is output six times input on OpenAI?
We do not know, and the page does not say. What we can confirm is that the multiplier holds on luna, terra and cyber, that sol sits at five times since its price cut, and that six times is the widest ratio among the major providers.
Are the older GPT models still available?
Yes, but not all of them for long. On 6 October 2026 their model pages still carry a price, and gpt-5.2 has no deprecation marker, while gpt-5.1, gpt-5, gpt-5-mini and gpt-5-nano are all marked deprecated. None of them appears any more on the main pricing page, whose flagship table now lists only the GPT-6 models.
What we verified and what we did not
Verified: input, cached input and output rates for gpt-5.6-luna, terra, sol and cyber, the one-tenth cache ratio across the family, the six-times output ratio on luna, terra and cyber and the five-times ratio on sol, the long-context rates for luna, terra and sol and their absence on cyber, and the continued listing of gpt-5.2, gpt-5-mini and gpt-5-nano without deprecation markers, all read from the OpenAI pricing page on 5 September 2026. The sol rates were $5.00 and $30.00 when this page was first published on 16 August 2026; OpenAI has since cut them to $4.00 and $20.00, and the long-context pair from $10.00 and $45.00 to $8.00 and $30.00.
Not verified, and therefore not asserted: the token threshold at which long-context pricing takes over, which the page does not publish; what separates the tiers in capability, which the pricing page does not address; rates on Azure, which publishes its own grid; and any benchmark claim, since this page prices models rather than ranking them.
One note on method. An earlier version of our OpenAI coverage was built on the top of this pricing page alone, and we briefly concluded that models further down it did not exist. They do. A long page has to be read to the bottom before anything is claimed to be missing from it, which is why this page lists what it checked as well as what it found.