GPT-5.6 is priced per token on the API, split across three tiers: Sol at $5 input / $30 output, Terra at $2.50 / $15, and Luna at $1 / $6, all per million tokens. If you use ChatGPT instead of the API, Sol appears as a reasoning setting on paid plans rather than a separate charge. The whole point of the three tiers is to stop you paying flagship rates for work a cheaper model handles just as well.
The short answer
For developers, Terra is the default — it matches GPT-5.5-class quality at roughly half the cost, so most production traffic should route there. Send only the hardest agentic and reasoning tasks to Sol, and push high-volume, simple work (classification, extraction, routing) to Luna. For everyday chat and writing, a flat ChatGPT subscription is usually the cheaper path than paying per token.
API pricing by tier
| Tier | Input / 1M | Output / 1M | Best for |
|---|---|---|---|
| Sol | $5.00 | $30.00 | Hard, long-running, agentic tasks |
| Terra | $2.50 | $15.00 | Balanced everyday production work |
| Luna | $1.00 | $6.00 | High-volume, simpler tasks |
Output tokens cost far more than input across every tier, which is the number to watch. A verbose agent that emits long chains of reasoning and tool calls will run up output costs fast — worth remembering before you point Sol at a high-volume workload.
Prompt caching — the discount to actually use
GPT-5.6 adds explicit prompt caching, and on the right workload it's the single biggest lever on your bill. If many requests share a large prefix — a system prompt, a knowledge base, a long instruction block — you cache it once and reuse it at a reduced input rate instead of paying full price every call.
For retrieval-augmented apps, agents with big system prompts, or anything that replays the same context repeatedly, caching can cut input costs substantially. It doesn't touch output pricing, so pair it with tight, well-scoped responses for the best effect.
Using GPT-5.6 in ChatGPT
If you're not building on the API, you reach GPT-5.6 through ChatGPT. Sol is exposed via reasoning settings on eligible paid plans, while Terra and Luna are aimed primarily at ChatGPT Work, Codex, and API users. That means the practical cost of Sol for most people is simply their ChatGPT subscription.
ChatGPT's paid tiers run from Plus at $20/month to Pro at $200/month, with higher usage limits and reasoning access as you go up. The ChatGPT pricing guide breaks down which plan unlocks what, so you don't pay for Pro when Plus would do.
Which tier should you use?
Route by difficulty, not by habit. A quick mental model: if a competent junior could do the task, send it to Luna; if it needs real judgment or code that must run, use Terra; if it's a genuinely hard, multi-step, autonomous problem, spend on Sol.
Most teams find that a small share of requests truly need the flagship, so a Terra-default, Luna-for-volume, Sol-for-hard-cases split keeps quality high while holding the bill down. That tiering is the main reason GPT-5.6 can be cheaper in production than a single-model setup, even with Sol's premium rates.
Is it worth it?
Yes, if you use the tiers deliberately. Terra alone is a strong value — GPT-5.5 quality at about half the price — and Luna makes high-volume automation cheap. Sol is worth its premium only for the hard problems where its benchmark lead actually shows up.
Before you commit an architecture to it, read the full GPT-5.6 review, and compare the frontier options in GPT-5.6 vs Claude Opus 4.8 and our best AI chatbots ranking.
Frequently Asked Questions
How much does GPT-5.6 cost in 2026?
On the API, Sol is $5 input and $30 output per million tokens, Terra is $2.50 and $15, and Luna is $1 and $6. In ChatGPT, Sol is available through reasoning settings on paid plans like Plus ($20) and Pro ($200).
Which GPT-5.6 tier is cheapest?
Luna, at $1 input and $6 output per million tokens, aimed at high-volume simpler tasks. Terra sits in the middle at $2.50/$15, and Sol is the flagship at $5/$30.
Does GPT-5.6 have prompt caching discounts?
Yes. Explicit prompt caching lets you reuse a cached prompt prefix at a reduced input rate. On repetitive workloads with large shared context, that can cut input costs substantially.
Why is output more expensive than input?
Generating tokens costs more compute than reading them, so every tier charges several times more for output. Verbose, reasoning-heavy responses drive most of the bill, which is why tight prompts and scoped outputs matter.
Is Terra really half the price of the old flagship?
Roughly, yes. OpenAI positions Terra at competitive performance to GPT-5.5 while being about 2x cheaper, which makes it the sensible default for most production traffic.
Do I need the API to use GPT-5.6?
No. Sol is available in ChatGPT through reasoning settings on eligible paid plans, so many users get it via their subscription. The API is for developers who want per-token access and tier control.