Not a Ladder: What OpenAI’s Own Numbers Say About Sol, Terra and Luna

The rule of thumb circulating since GPT-5.6 shipped goes something like this: start with Terra, move up to Sol when the task gets genuinely tricky, drop to Luna for high-volume simple work. Three tiers, one ladder, pick a rung.

It is a tidy mental model and I used it myself for a fortnight. Then I read OpenAI’s own published comparison tables properly, and the ladder fell apart. The numbers do not describe three points on a single axis of capability. They describe three different capability profiles that happen to be priced as though they were a ladder.

That distinction changes what you should actually do.

What the three tiers are, and what they now cost

GPT-5.6 reached general availability on 9 July 2026 across ChatGPT, Codex and the API. The number is the generation; Sol, Terra and Luna are what OpenAI calls durable capability tiers, each free to advance on its own cadence. That naming replaces the old mini and nano suffixes, and it is a better scheme, because it stops implying that the small model is merely a truncated version of the big one.

The pricing moved three weeks after launch, and a lot of published guidance has not caught up.

Work through the ratios, because they are the whole argument. Sol costs 25 times what Luna costs, on input and on output alike. Terra costs ten times Luna. At launch the top-to-bottom gap was five times; after the 30 July reduction it is 25 times. If your mental model of the price spread was formed in the first three weeks after launch, it is now wrong by a factor of five.

A 25x spread demands a much sharper routing policy than a 5x spread does. It also means the interesting question is not “is Sol better than Luna” — obviously it is — but “is Sol 25 times better at this specific thing”. Usually not. Occasionally, dramatically yes.

Where the tiers are effectively the same model

Here is the chart I wish someone had published on day one. It plots the gap between Terra and Luna across eighteen evaluations from OpenAI’s own tables, in percentage points.

Eleven of the eighteen land inside five points. Several are inside one point. On Agents’ Last Exam, the flagship measure of long-running professional workflows, Terra scores 50.4 and Luna 50.3. On GDPval, they sit at 1,593 and 1,592 Elo. On Toolathlon and BenchCAD, Luna is fractionally ahead.

Read that against the pricing table. On the benchmark OpenAI chose to lead with for professional work, the tier that costs ten times more wins by a tenth of a point.

That is not a small finding. It means that for a large class of real work — scoped implementation, tool-calling loops, classification, extraction, drafting, first-pass review, routine multi-step agent tasks — the default should not be Terra. It should be Luna, and the burden of proof should sit on escalation.

Where the cliffs are

The exceptions in that chart are not scattered. They cluster into four recognisable domains, and once you can name them the routing policy writes itself.

Long-context recall. This is the big one, and it is a genuine cliff rather than a gradient. On OpenAI’s MRCR v2 eight-needle test at 256K to 512K tokens, Sol scores 91.5 percent, Terra 89.6, and Luna 41.3. At 512K to 1M, Sol and Terra hold at 73.8 and 72.5 while Luna stays pinned at 41.3. GraphWalks at one million tokens tells the same story: 77.1, 71.2, 51.2. Anything that depends on finding and reconciling specific facts scattered through a large corpus — multi-document synthesis, whole-codebase reasoning, contract review across an entire data room — is the wrong job for Luna, regardless of how simple each individual step looks.

This is also precisely the claim I would push back on hardest in the popular summary of these models. They do not share the same effective long-context behaviour. The context window may accept the tokens; the recall does not survive them.

Security work. ExploitBench runs 73.5 percent on Sol, 52.9 on Terra, 33.2 on Luna. SEC-Bench Pro, which tests proof-of-concept generation against complex software, gives 71.2, 57.7 and 48.9. ExploitGym is starker still at 33.7, 23.2 and 12.4. For anyone doing defensive work — triage, patch validation, detection engineering — tier choice here is not a cost optimisation, it is a correctness decision. Note also that OpenAI gates the most capable cyber behaviour behind a verified-access programme, so the model you can reach may not be the model in the table.

Science. GeneBench Pro: 28.7, 23.3, 10.8. MedChemBench: 48.3, 35.0, 30.4. Luna loses more than half of Sol’s performance on genomics-style analysis. If you are building anything on top of life-science reasoning, this tier gap is larger than almost any prompt improvement you could make.

Low-level optimisation and hard mathematics. KernelGen 1P: 61.1, 49.2, 22.4. FrontierMath Tier 4: 83.0, 68.3, 58.5. ARC-AGI-3 is the most extreme entry in the whole table at 7.78, 0.8 and 0.18 percent — Sol is roughly ten times Terra and forty times Luna on abstract reasoning, which is the one place the ladder metaphor genuinely holds.

The revised rule

Default to Luna. Escalate on a named trigger, not on a feeling that the task is hard. Difficulty is a poor proxy, because the tiers are close on a great deal of difficult work and far apart on a few specific things.

In practice that means the highest-value engineering you can do here is not prompt tuning. It is building a router and an eval set. OpenAI’s own suggested pattern is worth copying: use Sol to resolve ambiguity and produce the plan, then hand well-specified implementation, test-writing and evaluation to Luna. The expensive model does the thinking that cannot be delegated; the cheap model does the volume.

Collect twenty to fifty real examples from your actual workload, including the edge cases that hurt. Run all three tiers with identical settings. Then look specifically at where the cheap model fails rather than at average scores, because an average hides the one failure mode that breaks production. Pick the cheapest tier that clears your quality bar, and re-check when prices move — which, as of 30 July, they demonstrably do.

Two oddities worth knowing

The tiers are not strictly ordered, even within OpenAI’s own numbers. On NanoGPT, Terra scores 14.5 percent against Sol’s 9.69. On PostTrainBench Lite, Terra takes 51.5 against Sol’s 50.3. On GraphWalks BFS at 256K, Luna beats Terra, 81.3 to 76.9. These are narrow evaluations and the differences are small, but they undercut any assumption that paying more is monotonically safer. Occasionally it is not.

Second, the levers other than tier are now substantial. Reasoning effort runs from medium through xhigh to max, and ultra coordinates four parallel agents by default. Prompt caching became more predictable in this generation, with explicit cache breakpoints, a thirty-minute minimum cache life, cache writes billed at 1.25 times the uncached input rate and cache reads still discounted 90 percent. For a long-running agent re-reading the same repository, caching will move your bill more than tier selection will. Fast mode, which replaced Priority Processing, buys up to 2.5 times the speed on Sol at twice the price with no change in intelligence, and OpenAI began previewing an Ultrafast mode this month claiming up to fourteen times.

Caveats I would not skip

Everything above comes from OpenAI’s published tables, and vendor benchmarks are vendor benchmarks. They pick the evaluations, the harnesses, the reasoning settings and the competitors’ configurations. Those tables also include scores for models from other labs, and the picture there is mixed rather than one-sided in any direction; read the source rather than anyone’s summary of it, including mine.

Several figures in OpenAI’s prose differ slightly from its own tables, which usually reflects a different reasoning setting rather than an error. I have used the table values throughout, since those are the ones with stated conditions.

And the numbers have a shelf life measured in weeks. This generation has already seen a price cut, a new speed tier and a preview of another inside five weeks of launch. Treat any routing policy, including this one, as something to re-derive each quarter from your own evals.

The short version

Sol, Terra and Luna are not three rungs. They are three profiles, priced 25 to one from top to bottom, that converge on most general agentic and professional work and diverge sharply on long-context recall, security, science and low-level optimisation.

So invert the usual advice. Start at Luna, not Terra. Escalate when the work crosses one of four named lines. And spend your effort on the eval set that tells you where those lines fall for your workload, because that artefact keeps paying every time the price list changes.