GPT-5.6 pricing: what Sol, Terra, and Luna actually cost

GPT-5.6 pricing at a glance

Here's the part you came for. These are the official preview rates from OpenAI's help-center article, per 1M tokens:

A couple of things worth flagging that the table alone won't tell you. There's no mini or nano variant for GPT-5.6 (that split lived on the older GPT-5.4 family), and OpenAI hasn't published a separate long-context input tier yet either. For GPT-5.5, going past roughly 272K tokens of context bumps the input rate; GPT-5.6's preview lists a single flat rate per model, which may change at general availability.

Sol, Terra, and Luna: what you're actually buying

GPT-5.6 introduces a new naming system, and it's worth understanding because it changes how you pick a model. The number (5.6) is the generation. Sol, Terra, and Luna are durable capability tiers that OpenAI says can advance on their own cadence, replacing the old mini/nano suffixes.

In plain terms:

• Sol is the flagship, the most capable model, with the highest price and the deepest reasoning. It's the one that sets a new state of the art on OpenAI's agentic coding benchmark.
• Terra is the everyday workhorse. OpenAI claims it has competitive performance to GPT-5.5 while being 2x cheaper, which makes it the default I'd start with for most production workloads, the way you'd pick a right-sized agent over a rule-based bot.
• Luna is the fastest and most cost-efficient, aimed at high-volume, latency-sensitive jobs where Sol-grade reasoning is overkill.

There are also two new compute controls that affect what you pay, because they change how many tokens the model burns. GPT-5.6 adds a max reasoning effort that gives Sol the most time to think, sitting on top of the existing low/medium/high levels. And it adds an ultra mode that fans work out across subagents, rather than running one longer chain of thought. On Terminal-Bench 2.1, Sol Ultra scores 91.9% versus 88.8% for plain Sol. The catch is that more reasoning effort means more output tokens, and output is the expensive half of the bill, so max and ultra are levers on your cost, not free upgrades.

How GPT-5.6 compares to GPT-5.5 and GPT-5.4

This is where the interesting story is, and it cuts against the pre-launch chatter. There were rumors OpenAI would cut frontier prices with this generation. It didn't. GPT-5.6 holds the line: Sol matches GPT-5.5's short-context rate exactly, and Terra lands on the old GPT-5.4 price point.

So the upgrade narrative is "more capability at the same price," not "same capability, cheaper." If you're already running GPT-5.5 at $5/$30, moving to Sol gets you a better model at no extra cost. The actual savings opportunity is Luna at $1/$6, a new low-cost frontier-family tier that sits between GPT-5.4 and GPT-5.4-mini. And there's no GPT-5.6 Pro yet, so the most expensive option in the lineup remains GPT-5.5 Pro at $30/$180.

It tracks with how the wider market is behaving. As one commenter put it on the Hacker News launch thread:

"AI is turning out to be a fairly competitive but 'normal' product. Companies carving out niches on cost, quality, and speed."

That's exactly what the three-tier split is: OpenAI carving Sol, Terra, and Luna into different cost-quality-speed corners rather than racing the headline price to zero.

What a real GPT-5.6 bill looks like

Per-million-token rates are abstract, so let me make it concrete. Say you're running a workload that uses 50M input tokens and 10M output tokens a month (a reasonable shape for an agent that reads a lot of context and writes shorter replies). Here's the monthly bill on each tier:

• Luna: (50 × $1) + (10 × $6) = $110/month
• Terra: (50 × $2.50) + (10 × $15) = $275/month
• Sol: (50 × $5) + (10 × $30) = $550/month

Same work, and the priciest tier costs 5x the cheapest. That gap is the whole reason the tiers exist: you route the hard requests to Sol and the rest to Luna.

One thing that can pull the bill down: prompt caching. GPT-5.6 introduces more predictable caching with explicit cache breakpoints and a 30-minute minimum cache life. Cache reads get the standard 90% discount (so cached input is roughly $0.50/1M on Sol), while cache writes are billed at 1.25x the uncached input rate. If your prompts share a big stable prefix (a long system prompt, a fixed knowledge base), caching is the single biggest lever on the input half of your bill.

The thing this worked example quietly assumes is that you know your token volume. On a coding agent you roughly do. On a customer support queue, you don't: token use per ticket swings with how chatty the customer is, how much history the model reads, and how many tool calls it makes. That's the part token pricing handles badly, and I'll come back to it.

ChatGPT pricing: where GPT-5.6 fits (for now, it doesn't)

If you reached this page expecting GPT-5.6 inside a ChatGPT subscription, here's the honest answer: it's not there yet. During the preview, the consumer ChatGPT tiers still serve GPT-5.5. For completeness, the current ChatGPT pricing looks like this:

On the business side, ChatGPT Business runs $20/user/month on annual billing ($25 monthly), and Enterprise is custom, contact-sales pricing with the expected SSO, data-residency, and SLA add-ons. None of these expose GPT-5.6 during the preview, so if you specifically need 5.6 today, the API and Codex are the only doors, and even those are gated.

The catch nobody puts in the headline: it's a gated preview

This is the most important line in the whole post if you're trying to plan a rollout. GPT-5.6 launched June 26, 2026 as a limited preview, available only through the API and Codex to a small group of vetted partners. There's no public waitlist or self-service signup; access goes through an OpenAI account representative. General availability is promised "in the coming weeks" with no firm date.

There's a second wrinkle that matters for anything customer-facing. OpenAI's own system card notes GPT-5.6 shows a greater tendency than GPT-5.5 to go beyond the user's intent, taking unrequested actions, though absolute rates stay low. The preview also ships with layered safety classifiers that can pause generation or slow responses in dual-use areas. For a coding sandbox that's fine. For an AI agent answering customers unsupervised, "occasionally does more than asked" is exactly the failure mode you build guardrails against. Having run frontier models on live support queues for a while now, I can tell you the model's raw benchmark score is rarely the thing that breaks, it's the unscoped action nobody simulated.

Token pricing vs paying per outcome

Here's the reframe I'd leave you with. GPT-5.6's pricing is great if you're a developer building on the API and your usage is predictable. But if your actual goal is "answer customer support tickets with AI," paying per token is the wrong unit, and I say that as someone who has watched teams try to forecast a support bill from token counts and give up.

Token billing counts every input and output token, so your bill moves with how long each conversation runs, how much knowledge the model reads, and how many tool calls it fires. Two tickets that both end "resolved" can cost wildly different amounts. That unpredictability is the single most common reason I hear teams stall on rolling out AI support: they can't put a number in next year's budget, which is the whole point of tracking AI support cost savings against a unit you control.

This is the same build-versus-buy fork a lot of teams hit. As Karel from GENERAL BYTES put it in a build-versus-buy conversation:

"We could try to write our own LLM application but we didn't want to invest our time into that. We wanted something that we would not have to maintain."

That's the trade. You can wire GPT-5.6 to your helpdesk, manage the token spend, build the guardrails, and maintain it. Or you can pay per resolved ticket and let someone else own the model layer, the hallucination controls, and the upgrade path when GPT-5.7 lands.

Try eesel

If your reason for reading a GPT-5.6 pricing page is customer support, this is the part that's actually relevant to you. eesel is an AI support agent that runs on frontier models like the GPT-5.x family, but you never touch a token meter: it's priced per resolved ticket, so the cost scales with outcomes you can actually forecast, not with how verbose a conversation got. It plugs into your existing helpdesk, learns from your past tickets and knowledge base, and you can simulate it against your real ticket history before it ever replies to a customer, so you see the resolution rate (and the cost) up front. It's free to try, and you don't need to be on an OpenAI preview list to start.