OpenAI split GPT-5.6 into three tiers and priced the flagship exactly where GPT-5.5 already sat. The tier worth your attention is the cheaper one in the middle.
Sol, Terra, and Luna arrived on June 26 at $5/$30, $2.50/$15, and $1/$6 per million tokens. Sol, the flagship, costs to the cent what GPT-5.5 costs, so the headline tier is not where the money moves. Terra is: GPT-5.5-class work at half the rate. And there is a quieter line in the fine print, where Sol's ultra mode keeps the same price per token while spending a lot more of them.
Photo by Benjamin Voros on Unsplash
One thing to set straight before any table: you almost certainly cannot call GPT-5.6 yet. It went out on June 26 to about 20 partner organizations as a limited preview, and OpenAI's own pricing and preview pages sit behind a bot wall we could not read directly. The numbers below are the ones every pricing tracker is quoting from OpenAI, cross-checked across several of them and consistent to the cent. Read them as provisional preview rates, not a settled public price sheet.
Three tiers, three rates
The family drops the decimal-point naming for once and reaches for the sky. Sol is the sun, the flagship. Terra is earth, the everyday tier. Luna is the moon, small and fast. Capability and price fall in that order.
| Tier | Input | Output | Cached input | Meant for |
|---|---|---|---|---|
| Sol | $5 | $30 | $0.50 | Hardest agentic + coding |
| Terra | $2.50 | $15 | ~$0.25 | Balanced production default |
| Luna | $1 | $6 | ~$0.10 | Fast, routine, cost-sensitive |
Per million tokens. Cached reads take the standard 90% discount, so Sol's $0.50 is confirmed; the Terra and Luna cached figures are the same 90% math applied to their input rates and were not published separately, so we mark them approximate. Cache writes bill at 1.25x the input rate with a 30-minute minimum life.
The flagship is a price hold, not a price cut
Put Sol next to the model it replaces and nothing moves on the sticker. GPT-5.5 is $5 in and $30 out. Sol is $5 in and $30 out. OpenAI is asking the same money for what it says is a better model, which is a defensible move and a boring one for anyone whose job is watching the meter. If you were already paying GPT-5.5 rates, Sol changes your capability, not your budget line.
That output rate is worth staring at for a second. At $30 per million, Sol is the most expensive output on our pricing page among the general-purpose flagships. It sits $5 above Claude Opus 4.8, which does output at $25, and it towers over anything from the open-weight camp. The pitch has to be that Sol earns the premium on the hardest work. Which brings us to the number OpenAI actually led with.
On Terminal-Bench 2.1, an agentic coding benchmark, Sol posts 88.8% in its base setting and 91.9% with ultra mode on. Those two figures are the cleanest data in the whole launch, agreed on across every source. Almost everything else OpenAI cited came without a comparable number, and it did not publish a SWE-bench Verified score at all, so we are not going to invent one. The one benchmark it stood behind says Sol is genuinely good at running a terminal. Whether that is worth $30 output over Opus 4.8's $25 depends entirely on the task in front of you.
Terra is where the interesting math lives
Here is the tier OpenAI buried in the middle of the announcement and probably should have led with. Terra runs $2.50 in and $15 out, which is Sol cut straight down the middle and the same 50% off what GPT-5.5 charged right up until this launch. OpenAI describes Terra as competitive with GPT-5.5, so take that claim even loosely and you are looking at last generation's flagship intelligence for fifty cents on the dollar. That is the kind of move that actually shifts what people run in production.
It also lands right on top of a crowded fight. Terra's $2.50/$15 undercuts Claude Sonnet 5's standard $3/$15 on input and matches it on output, and it sits well under Opus 4.8's $5/$25. Suddenly the balanced tier from three different labs clusters within a couple of dollars per million. The differentiators stop being the rate card and start being the tokenizer, the caching behavior, and which model your existing code already talks to.
The line in the fine print: ultra mode is free per token and expensive per task
Sol carries two reasoning settings the cheaper tiers do not. Max deepens a single reasoning chain. Ultra goes wider, spinning up parallel subagents that each chew through the problem and report back. Ultra is what pushed Terminal-Bench from 88.8% to 91.9%. Neither mode has its own price. Both bill at Sol's $5/$30.
That is the trap, and it is an easy one to walk into. A rate that does not change reads like a cost that does not change. But parallel subagents generate parallel token streams, and you pay for every one of them at $30 per million output. The per-token price holds; the token count does not. Flip on ultra for a task and you can move the same rate card a long way up the invoice without a single number on the pricing page changing.
OpenAI has not published how many extra tokens ultra spends, so we will not put a multiplier on it. The rule to carry is simpler: on Sol, the tier you pick sets the rate, but the mode you pick sets the bill. Reach for ultra when the extra three points on a benchmark are worth an unknown but real multiple of the token spend, and leave it off for everything routine.
What each tier does to a monthly bill
Rates per million are hard to feel. So here is a heavier coding-agent shape, 50M input and 10M output in a month with no caching, run through each GPT-5.6 tier and the models they compete with.
| Model | Input | Output | 50M in / 10M out |
|---|---|---|---|
| GPT-5.6 Luna | $1 | $6 | $110 |
| Claude Sonnet 5 (intro) | $2 | $10 | $200 |
| GPT-5.6 Terra | $2.50 | $15 | $275 |
| Claude Sonnet 5 (standard) | $3 | $15 | $300 |
| Claude Opus 4.8 | $5 | $25 | $500 |
| GPT-5.6 Sol = GPT-5.5 | $5 | $30 | $550 |
All list rates, per million tokens, monthly column rounded to the dollar at 50M input and 10M output with no caching and no reasoning-mode surcharge. Sonnet 5 intro pricing runs through Aug 31, 2026. GPT-5.6 figures are preview rates.
The spread is the whole point. Luna does that month for $110, cheaper than even Sonnet 5's intro price and a fifth of what Sol charges for the identical traffic. Terra's $275 slides just under Sonnet 5's standard rate and roughly halves Opus 4.8. Sol, at $550, is the price of conviction: you pay it because the task genuinely needs the top of the range, not because it is a bargain. And none of these include ultra mode, which would push the Sol line higher by an amount OpenAI has not told anyone.
You still cannot buy any of it
A price sheet you cannot act on is a strange thing to publish, and yet here we are. GPT-5.6 is a closed preview for roughly 20 organizations, reachable through the API and Codex, as VentureBeat first reported. The staggered rollout traces to a June 2, 2026 US executive order that has federal agencies building a review process for frontier models, so OpenAI shared the models and its release plans with the government before opening the doors. We wrote about that gating and the models you can still call in a separate piece last week.
The one concrete promise on speed is Cerebras. OpenAI says Sol will run there at up to 750 tokens per second starting in July, initially for select customers as capacity comes online. Broad availability is the usual coming weeks, with no date attached. Until then, the pricing above is something to plan around rather than spend against.
If you need frontier-class output today, the field is already open. Opus 4.8 and Sonnet 5 are live, GPT-5.5 is unchanged at the same $5/$30 Sol will inherit, and the cheaper open-weight coders sit an order of magnitude below all of them. Drop your own token mix into the calculator before you pencil GPT-5.6 into a budget, because the tier you end up needing, and the mode you leave switched on, will decide the bill far more than the launch-day headline did.
Sources
- - OpenAI GPT-5.6 preview announcement (official, behind bot wall at time of writing): openai.com
- - VentureBeat (availability, US government review, Cerebras): venturebeat.com
- - MarkTechPost (tiers, reasoning modes, Terminal-Bench): marktechpost.com
- - eesel AI (pricing detail, long-context note): eesel.ai
- - Finout (cached, batch, residency pricing): finout.io
- - Eden AI (benchmarks, capabilities): edenai.co
- - TokenCost pricing page: tokencost.app/pricing
OpenAI's own pricing and preview pages returned a bot challenge at the time of writing, so every figure here is drawn from independent trackers quoting OpenAI and cross-checked between them. Context window, max output, tokenizer, and any long-context surcharge for GPT-5.6 were not published; we have left them out rather than guess. We will update this post against the official page once the model reaches general availability.