GPT-5.6 Sol vs Terra vs Luna: which model to use

OpenAI shipped GPT-5.6 today in three tiers, and within a week most teams will make the same mistake: they'll point everything at Sol, the flagship, because it tops the benchmark charts, and they'll overpay by a factor of five for work that

OpenAI shipped GPT-5.6 today in three tiers, and within a week most teams will make the same mistake: they'll point everything at Sol, the flagship, because it tops the benchmark charts, and they'll overpay by a factor of five for work that a cheaper model does just as well. The right way to use GPT-5.6 is to match the model to the job: Sol only for genuinely hard problems, Terra as the everyday default, and Luna for high-volume simple work, which means the tier that tops the leaderboard is the wrong choice for most of what you actually run. There's also one genuine surprise in the benchmarks worth knowing before you pick.

Here's the tier-by-tier breakdown, the real pricing, and how we decide on a client build.

The three tiers, plainly

Sol is the flagship: the strongest, most capable model in the family. It's the only tier that unlocks the new "max" reasoning effort and Ultra Mode, where it coordinates subagents instead of working alone. Best for hard, multi-step problems where you want maximum capability and a wrong answer is costly.

Terra is the balanced workhorse. OpenAI positions it as competitive with the previous GPT-5.5 while costing about half as much. This is the everyday default for most real work: drafting, summarizing, classification, routine reasoning.

Luna is the cheap, fast tier. Built for high volume and latency-sensitive tasks, it's the one you reach for when you're running millions of tokens through a simple, repetitive job and speed and cost matter more than frontier intelligence.

The pricing, side by side

Price is where the tiering decision actually gets made, so start here.

That five-times gap between Sol and Luna is the whole game. On a low-volume task you run a few hundred times a month, the absolute difference is pocket change and you may as well use the best model. On a high-volume automation pushing millions of tokens, that same gap decides whether the workflow is profitable. The tokens are the running cost of the automation, a point we hammer in the €500/month AI stack for a small team: model choice is a line item you tune, not a badge you wear.

The benchmark surprise: cheaper is not always weaker

Here's the non-obvious part. On Terminal-Bench 2.1, a coding and command-line benchmark, the cheapest tier is not the weakest.

Model	Terminal-Bench 2.1
Sol Ultra	91.9%
Sol	88.8%
GPT-5.5 (prior gen)	88.0%
Luna	84.3%
Terra	82.5%

Read that table twice. Luna, the cheapest model, scores higher than Terra on this particular benchmark. That doesn't make Luna better overall (Terra wins on other tasks, and benchmarks are narrow), but it kills the lazy assumption that price equals capability in a straight line. The lesson: don't pick a tier by its position in the price list. Pick it by how it does on the kind of work you'll actually run. A single benchmark is a data point, not a verdict, which is the same caution we apply in our honest coding-CLI comparison.

Which model for which job

Strip away the charts and the decision is short.

Most real automations end up using more than one. A pipeline might route and classify with Luna, draft with Terra, and hand the one genuinely hard step to Sol. Paying flagship prices for the whole chain is the overspend we see most often.

Where Sol is worth it, and where it's a waste

Sol is a real step up on the hardest problems, especially in Ultra Mode, where it coordinates subagents and posts the top benchmark score. If your task is a long, brittle, multi-step chain where reliability is worth real money, Sol earns its keep, and we cover why in what Ultra Mode's subagents change for automation.

Where it's a waste: everyday drafting, simple classification, high-volume anything. Using Sol to write a confirmation email is like renting a race car to get groceries. It'll do it, and you'll have paid five times too much for the privilege. If the version-number noise has you unsure whether any of this even applies to your business, we wrote the plain-English take in does GPT-5.6 change anything for your business.

The competitor question, answered honestly

You can't talk about GPT-5.6 without the obvious question: what about Claude? At launch, Sol leads the coding benchmarks against Anthropic's frontier Fable Mythos line. But leading a benchmark this week is not the same as being the right tool for your job.

We build with OpenAI and Anthropic both, and we pick per workload, not per team. Claude still wins plenty of writing and reasoning tasks; GPT-5.6 wins others and just got cheaper. Anyone telling you one model is the answer for everything is selling a preference, not an assessment.

— how we actually decide, and it isn't loyalty

The honest position is boring: the frontier is close, it changes every few months, and the right model is the one that wins your specific job on your specific data. We keep the same posture on the build-versus-orchestrate question in n8n vs Claude agents. Loyalty to a vendor is how you overpay.

How we'd pick for a client build

On an actual project we don't start with the model. We start with the job, its volume, and its tolerance for a wrong answer. High volume plus low stakes goes to Luna. Everyday work goes to Terra. The one hard step that can't fail goes to Sol, often in Ultra Mode. Then we measure the cost per useful output and adjust. The model is a dial we tune against a number, not a flag we plant. Picked that way, GPT-5.6 is mostly good news: the same automation you'd have built last week now has a cheaper, steadier engine under most of it.