The Vibe Father
Pricing decoded
OpenAI shipped the GPT-5.6 family on July 9, 2026 with three prices that tell you exactly how OpenAI wants you to use it: Sol at $5 in / $30 out, Terra at $2.50 / $15, and Luna at $1 / $6. Read those numbers as a job assignment, not a menu. The tiers are not "good, better, best" — they are "thinker, builder, scout," and the whole point of a family is that you run all three, each on the work its price is designed for.
We run a live coding leaderboard, and the question we get most is never "which model is best" — it is "which model do I run for this, and what will it cost." So let's decode the family by workload, then do the arithmetic on a real month.
The three prices, and what each is for
Sol ($5 / $30) is the planner. Output tokens are where Sol is expensive, and a planner emits relatively few of them — a plan, a design, a diagnosis — while its decisions steer everything the cheaper models do next. You run the planner rarely and it earns its rate on leverage, not volume. Cheap in tokens for the value it returns.
Terra ($2.50 / $15) is the volume builder. Exactly half of Sol on both sides, Terra is the model you keep busy turning an approved plan into diffs across many turns. This is where output-token pricing compounds, so the half-rate versus Sol is the point of the tier.
Luna ($1 / $6) is the scout. A fifth of Sol's rates. Cheap enough to run constantly for reconnaissance, triage, and low-ambiguity volume edits — the work you do a lot of and grade with a test gate rather than trust.
The heavy-month math
Our standard yardstick is a heavy month of 50M input and 10M output tokens. Run the whole month on a single model and the arithmetic is simple — multiply input millions by the input rate, output millions by the output rate:
| Model | In / M | Out / M | 50M in + 10M out |
|---|---|---|---|
 GPT-5.6 Sol | $5 | $30 | $550 |
| GPT-5.6 Terra | $2.50 | $15 | $275 |
| GPT-5.6 Luna | $1 | $6 | $110 |
 Claude Opus 4.8 | $5 | $25 | $500 |
| Claude Sonnet 5 | $3 | $30 | $300 |
So Sol lands at $550 for that month — a hair above Claude Opus 4.8's $500 and well above Sonnet 5's $300. Terra comes in at $275, just under Sonnet. Luna at $110 is the cheapest of the lot by a wide margin. Those are the single-model figures, and they are useful for one thing: setting the ceiling.
Nobody runs one model for everything
Here is the mistake the table above invites: nobody sane runs their whole month on one tier. If your entire 50M/10M lived on Sol, you would be paying planner prices to do scout work, and $550 would be a self-inflicted wound. The family exists so you don't do that.
A realistic split routes the heavy thinking to Sol, the bulk building to Terra, and the volume reconnaissance to Luna. Say a quarter of your output tokens are genuine planning (Sol), half are building (Terra), and a quarter are scouting (Luna), with input distributed similarly. The blended month lands well under any single-model figure above — you are paying $30-output rates only for the tokens that actually earn them. That is the discipline: match the token to the tier. We work through this style of per-workload accounting throughout the economics of AI coding.
The efficiency wildcard
One number could rewrite this entire guide, and it is OpenAI's claim, not our measurement. OpenAI reports that Sol reaches its Terminal-Bench 2.1 result using less than half the output tokens and less than half the time of Claude Fable 5, at about a third less cost per answer. If that token efficiency generalizes to your real workloads — and that is a genuine "if," pending independent verification — then Sol's $30 output rate is misleading in your favor, because you would be buying fewer of those expensive tokens per finished task.
Per-token price is never the whole cost. A terse model at a high rate can be cheaper end-to-end than a chatty model at a low one, because what you actually pay for is tokens-to-done, not the sticker. Until independent evals confirm Sol's efficiency claim, treat it as a reason to run the math on your tasks rather than a reason to trust ours.
How to actually decide
Do not pick a GPT-5.6 tier from this table. Pick from your own workload. Take one representative task of each kind — a hard planning problem, a routine build, a batch of scouting — run each through the tier we suggest behind a verification gate, and record the end-to-end cost: tokens burned, retries needed, review time spent. The tier that finishes your task for the least total money wins that seat, full stop. That is the same measure-don't-guess approach we apply in our cheapest coding models guide.
Because the Vibe Father runs 22 CLIs and every model side by side on one macOS command deck, this routing is a per-agent dropdown, not three separate integrations. You put Sol in the planner seat, Terra on the builders, and Luna on the scouts, and the harness bills each to the model you assigned — the day the family's API opens, with no tool switch. When independent coding numbers for Sol, Terra, and Luna publish, they will land on the leaderboard, and you can re-slot any seat that a rival wins on performance-per-dollar. Until then: the prices are confirmed, the routing logic is sound, and the efficiency claim is OpenAI's to prove.
