OpenAI shipped GPT-5.6 on July 9 in three tiers — Sol, Terra, and Luna — and the single most expensive mistake you can make with it is treating that as a quality ladder where you always climb to the top. It's a routing menu. The tiers span a 5x price gap, and most of the work a founder runs through an API doesn't need the flagship.
If you read one line: default to Terra, drop to Luna for high-volume cheap work, and escalate to Sol only when a cheaper tier visibly fails.
The prices, plainly#
Per 1 million tokens:
- Luna — $1 in / $6 out. Fastest and cheapest.
- Terra — $2.50 in / $15 out. Balanced; roughly half of Sol.
- Sol — $5 in / $30 out. Flagship. There's also a Sol Ultra high-effort mode for the hardest jobs.
That's a 5x spread on both input and output between Luna and Sol. Send a bulk classification job — thousands of short calls — to Sol instead of Luna and you've quietly multiplied that line item by five for output you probably couldn't tell apart.
One number matters more than the tier you pick: cache reads get a 90% discount (cache writes cost 1.25x). If your prompts share a big stable prefix — a system prompt, a knowledge base, a long instruction block — structuring them so that prefix is cached is a bigger lever on your bill than any tier choice. Do that first.
They're closer than the price gap suggests#
The reason routing works is that the cheap tiers aren't far behind on a lot of real tasks. On Terminal-Bench 2.1 (agentic coding), OpenAI's launch benchmarks report:
- Sol Ultra — 91.9%
- Sol — 88.8%
- Terra — 84.3%
- Luna — 82.5%
For hard, multi-step coding, that 6-point gap between Luna and Sol is real and worth paying for. But most production calls aren't Terminal-Bench. For classification, extraction, routing, and summarization, the gap between tiers is usually invisible in the output — which is exactly why paying flagship rates for them is waste.
(Treat cross-vendor benchmark tables floating around launch week with caution; the internally consistent, useful signal is the spread between the three GPT-5.6 tiers themselves, from OpenAI's own numbers.)
How to route, by task#
- Luna → anything high-volume and latency-sensitive: tagging support tickets, extracting fields from documents, routing intents, first-pass summaries of a large corpus. Cheap, fast, and good enough.
- Terra → your everyday default: chat and assistant features, content drafting, most steps inside an agent loop, RAG answers. This is what runs when you don't have a reason to go cheaper or more expensive.
- Sol / Sol Ultra → the hard tail: genuine reasoning, planning, agentic coding, anything where you can see a cheaper tier getting it wrong. Reserve max-effort Sol Ultra for the jobs where the extra latency clearly earns its keep.
The context window doesn't enter into it: all three tiers share the same ~1,050,000-token context and 128K max output. Tier is a cost-and-reasoning decision, not a "how much fits" decision.
The founder move: a tiny router, not a default#
You don't need an ML routing model. You need a two-line policy in your code:
- Start every new feature on Terra.
- For any endpoint that's high-volume and simple, try Luna and diff the outputs against Terra on a sample. If you can't tell them apart, ship Luna.
- For any endpoint where quality visibly breaks, bump that endpoint — and only that endpoint — to Sol.
Reassess when your volume grows: the endpoint that didn't matter at 1,000 calls a day is your biggest bill at 100,000. A per-endpoint tier map that you revisit monthly captures nearly all the savings a fancy router would, with none of the complexity.
The takeaway#
GPT-5.6's three tiers aren't good/better/best — they're cheap/default/heavy, and the win is matching each job to the smallest tier that still does it. Cache your stable prompt prefixes, default to Terra, push volume to Luna, and spend Sol only where it earns the 5x. That discipline is the difference between an AI bill that scales with revenue and one that scales with traffic.
Deciding which agent front-end to run GPT-5.6 behind? See Claude Cowork vs ChatGPT Work. Want the model to reliably repeat your workflow? Build an agent skill.
