GPT-5.6 Sol vs Terra vs Luna: Which Tier Do You Actually Need?

GPT-5.6 shipped on July 9 in three tiers. OpenAI's list pricing, per million input/output tokens:

The obvious question is which one you actually need. The obvious answer — "start at Luna, move up when it breaks" — is right, but the arithmetic behind it isn't what the price table implies.

We ran all three through the same battery. Every task grades by script against a known answer.

All three tiers passed everything

That includes the debug task, which is not trivial — the function has an input-mutation bug and a wrong even-length case, and a pass requires fixing both. Luna, at a fifth of Sol's list price, found both.

We tried to find the tier where cheap stops working, on tasks representative of ordinary production use. On this battery it doesn't exist.

The 5× discount is really 3.3×

Here's the part that doesn't show up in the pricing table. Summed across the three tasks:

Sol costs 3.34× what Luna does — not the 5× the list price implies. Luna's per-token rate is exactly 5× lower on both input and output. The gap closes because Luna emits 55% more output tokens for the same work, and it does that consistently:

Same verdicts, more words to reach them. Terra tracks Sol closely on verbosity, so its measured ratio (2.11×) lands almost exactly on its list ratio (2×). Luna is where the sticker price and the invoice diverge.

This matters when you're sizing a budget from a pricing page. A "5× cheaper" tier that talks 55% more is a 3.3× saving. Still a large saving — just not the one you planned for.

The cheapest tier is not the fastest

Terra beat Luna on every single task despite costing 2.5× more per token. Same reason: tokens take time to generate, and Luna generates more of them. Luna's cost advantage is real, but it buys you a slower response than the mid tier.

If you picked Luna expecting the budget tier to also be the fast one — a reasonable assumption, and wrong here — Terra is the tier you actually wanted.

Where they do differ: hard constraints

The three tasks above have correct answers the models can reason toward. Output constraints are different: they require suppressing tokens, not finding them. We ran "three sentences about the ocean, no letter e" five times per tier:

There's a ranking here, and it goes the direction you'd expect. But at five trials the gap between 3/5 and 4/5 is one coin flip, and we won't claim it's real. What the data does support: none of these tiers reliably honors a hard output constraint. The best of them misses one run in five.

For context, in yesterday's head-to-head Sol went 5/5 on this same test. Across both runs that's 9 of 10 — good, not dependable. If your product depends on a hard rule, validate the output; don't trust the tier.

How to pick

Start at Luna. It passed every ordinary task we ran, at a third of Sol's real cost. For extraction, classification, summarization, and routine code work, the evidence here says the budget tier is sufficient.

Move to Terra when latency matters. It's the fastest tier, and its list price is honest about what you'll pay because it isn't verbose. For interactive and user-facing paths, Terra is the sweet spot — not Luna.

Reserve Sol for genuinely hard problems. Nothing in this battery justified it. That's not a knock on Sol; it means our battery isn't hard enough to need it, and most production prompts aren't either. Sol earns its price on frontier-difficulty work — which is exactly what OpenAI's own benchmarks measure and exactly what routine pipelines aren't.

Budget from measured cost, not the price table. The list ratio told us to expect 5×. We measured 3.3×. Run your own prompts and read the actual token counts before committing a forecast.

Run it yourself

No account or API key needed:

• The extraction task — the one all three passed
• The constraint test — the one none of them reliably passes

To compare tiers side by side on your own prompts and see the token counts that drive the bill, that's what Promptster is for.

Run 2026-08-04 against gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna via the Promptster API, max_tokens: 4000, default temperature. Costs computed from list pricing (Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per million input/output tokens). The three graded tasks were run once each — enough for pass/fail on a deterministic answer, not enough to characterise latency variance, so treat timings as indicative. The constraint task was run five times per tier. Latency includes network round-trip from a single region.