GPT-5.6 Luna vs Sol: Tiers Instead of a Dial
OpenAI stopped shipping one model with an effort setting and started shipping separate tiers. Luna and Sol sit at opposite ends — here is how to decide which one a task needs.
The most consequential thing about GPT-5.6 is structural. Previous generations shipped as one model with an effort setting you could turn up. This one shipped as distinct tiers, each tuned to a different point on the cost–speed–capability curve.
That change moves a decision from runtime to design time. You no longer nudge a dial mid-task; you pick a tier up front and live with it. FrameTide carries two of them: Luna at 20 credits and Sol at 40.
The two tiers
| GPT-5.6 Luna | GPT-5.6 Sol | |
|---|---|---|
| Role | fastest, cheapest | flagship |
| Launch price | $1 / $6 per Mtok | $5 / $30 per Mtok |
| After July 30 cut | $0.20 / $1.20 | unchanged |
| FrameTide credits | 20 | 40 |
| Terminal-Bench 2.1 | — | 88.8% (Sol Ultra: 91.9%) |
Sol was previewed on June 26, 2026 and released publicly on July 9. OpenAI also ships a Terra tier between the two, and a Sol Ultra high-effort mode above it; FrameTide exposes Luna and Sol.
The July 30 price cut is the detail worth noticing: OpenAI dropped Luna by 80% while leaving Sol untouched. That is not a rounding adjustment, it is a statement about where they expect volume to go — the cheap tier is meant to absorb most traffic, and the flagship is meant to be reached for.
What Sol is actually for
Sol's strongest published result is agentic coding. At 88.8% on Terminal-Bench 2.1 it edges the frontier models it competes with, and Sol Ultra pushes that to 91.9%.
Terminal-Bench measures something specific: multi-step work in a real terminal, where the model has to run commands, read what came back, and recover from its own mistakes. That is a fair proxy for the jobs where a stronger model genuinely pays — long tool-call chains, migrations, anything where an early wrong turn compounds instead of just producing a bad paragraph.
Reach for Sol when the task has depth: many dependent steps, a wrong turn that is expensive to unwind, or a problem that has already defeated something cheaper.
What Luna is actually for
Almost everything else. Drafting, rewriting, summarising, answering questions, first-pass code, formatting, extraction — the work that makes up most of a day.
At 20 credits Luna is not the cheapest option in FrameTide. DeepSeek V4 Flash costs 1 and Gemini 3.6 Flash costs 5. So the honest framing is: Luna is the cheap tier of OpenAI, not the cheap tier of the lineup. You choose it when you specifically want an OpenAI model — a house style you have tuned prompts around, a behaviour your workflow depends on — rather than when you want the lowest price.
If you have no particular attachment to OpenAI's voice, the 1-credit and 8-credit models will do the same routine work for a fraction of the cost.
Choosing between them
The tier split makes this easier than it used to be, because it forces the question up front.
Start on Luna. Nearly every task begins as something a fast tier can handle. You find out otherwise quickly and cheaply.
Escalate on evidence. Move to Sol when Luna has actually failed at this task — not because the task feels important. Importance and difficulty are different axes, and only difficulty is worth 2x.
Escalate for depth, not length. A long document is not a hard problem; feed it to something with a big context window. A twenty-step chain where step three determines step nineteen is a hard problem, and that is Sol's territory.
Do not use Sol as insurance. The failure mode of tiered pricing is treating the flagship as the safe default. It is safe for the work that needs depth and expensive for everything else, and "everything else" is most of the work.
Against the rest of the lineup
- vs Claude Sonnet 5 (50). Sonnet is generally the stronger writer; Sol has the better published agentic-coding number. Route prose to Anthropic and long tool chains to Sol.
- vs Grok 4.5 (40). Same credit cost, similar purpose — Grok is built explicitly for long agentic runs and trained on real editor sessions, but its context dropped to 500K. Sol is the more general choice.
- vs Gemini 3.6 Flash (5). Four times cheaper than Luna with a 1M context. For anything input-heavy, Flash reads more for far less.
- vs DeepSeek V4 Flash (1). Twenty times cheaper than Luna. If you are using Luna for routine work out of habit rather than preference, this is where the money is going.
The short version
Luna is OpenAI's default and Sol is OpenAI's escalation. Pick the family first — if you want an OpenAI model, start on Luna and move to Sol when depth demands it. If you are choosing purely on capability per credit, look at the cheaper tiers first; the 20-credit floor here buys you a specific house style, not the best price in the lineup.
Sources
Create with GPT-5.6 Luna
Open FrameTide Agent with GPT-5.6 Luna already selected, and use it on your own creative task.