The Smartest Model Is the One You Didn’t Use

Today OpenAI previewed GPT-5.6, and for once the most interesting thing is not a benchmark. It is the menu. The release arrives as three named models: Sol, the flagship built for the hardest problems like complex coding and security research; Terra, a mid-tier model aimed at high-volume business work such as customer support, internal tools, and document analysis; and Luna, a fast and inexpensive model for everyday tasks like summarization, drafting, and routine automation. Access is limited at launch to roughly twenty organizations, after OpenAI shared the models and its release plans with the U.S. government, with a general release promised in the coming weeks.

Look at how that is organized. The number, 5.6, marks the generation. The names mark something more durable: tiers of capability that OpenAI says can advance on their own cadence. In other words, the company did not ship one model and ask you to take it or leave it. It shipped a price list and asked you a question you now have to answer for every task you run: how much intelligence does this actually need?

A six times spread from the cheapest output to the most expensive, inside a single product line, named after the sun, the earth, and the moon. See current rates for hundreds of models on our AI model pricing page.

There is a quiet admission buried in those three names. You do not send the sun to do the moon’s job. A frontier lab has stopped pretending that the goal is to put its biggest model in front of every request. The goal, it now says out loud, is to match the body to the task. That is a pricing strategy, yes. It is also a philosophy, and it happens to be ours.

01 The race quietly changed lanes

For three years the competition between AI labs was a contest of ceilings. Whose model scored highest, reasoned deepest, topped the leaderboard this month. Capability was the headline and price was a footnote you found later in the documentation. GPT-5.6 inverts that. The capability story is real, but the structure of the announcement leads with cost tiers. The first decision OpenAI wants its customers to make is not which generation to use. It is which budget to spend.

That is the signal worth paying attention to. When the most watched lab in the field organizes its flagship release around how much a task is worth rather than how clever the model can be, the center of gravity has moved. The frontier is no longer only about what is possible. It is increasingly about what is worth paying for.

When intelligence is the scarce thing, you buy the most you can. When intelligence is cheap and everywhere, the scarce thing becomes knowing where to spend it.

02 The economy of models

At superpenguin we have spent a long time treating models as an economy rather than a trophy case. The framing is simple to say and hard to live by: in the world of AI, the real limitation is your budget, and the advantage goes to whoever allocates it well. Capability is necessary, but capability without allocation is just an expensive way to do cheap work.

The instinct most teams still carry is to reach for the strongest model available and point it at everything. It feels safe. It is also, very often, the wrong call. A large share of real work is mechanical. It is the drafting, the summarizing, the reformatting, the routine automation that a smaller model handles at a fraction of the cost with no meaningful drop in quality. Spend a flagship on it and you have not bought better results. You have bought a bigger invoice.

The number that matters is not the price on the sticker and it is not the score on the leaderboard. It is the cost of getting a task done to an acceptable standard. Measured that way, the strongest model is frequently the least efficient choice, and the smartest move is the one nobody celebrates: the task you quietly handed to a cheaper model that did it just as well. Hence the title of this piece. The smartest model is often the one you did not use.

03 The floor is falling out from under the price

None of this would matter much if a token still cost what it cost two years ago. It does not, and the reason allocation is becoming urgent now is that the cost of intelligence is collapsing on two fronts at once: the hardware underneath it, and the chips being built specifically to run it.

Start with Nvidia. Its Vera Rubin platform, unveiled at CES in January and ramping into production in the second half of this year, is advertised at up to ten times lower cost per token and roughly five times the inference performance of the Blackwell generation it replaces, while needing about a quarter of the GPUs to train a comparable model. Rubin is not one chip but a co-designed system of six, including a part purpose-built for the long-context inference that agents and coding tools lean on. When the platform that most of the industry runs on cuts cost per token by an order of magnitude, every model that rides on it gets cheaper to serve.

Then look at what OpenAI did two days before this model launch. Together with Broadcom it unveiled Jalapeno, its first custom chip, an inference processor designed from scratch for large language models and brought from concept to silicon in about nine months. It is an ASIC, less flexible than a general purpose GPU but cheaper to run and tuned for exactly one job, and OpenAI says early testing shows performance per watt substantially better than the current state of the art. First deployments are aimed at the end of this year as part of a far larger plan to stand up its own accelerators at scale.

Jalapeno almost certainly is not what made Luna cost a dollar today. The timing is too tight and the chip is barely out of the lab. But it is the clearest possible statement of intent. A company does not spend nine months and a partnership with Broadcom building custom inference silicon unless it believes the future is won on cost per token, not just on capability. Read Vera Rubin and Jalapeno together and the trajectory is unmistakable: the floor under the price of intelligence is dropping, and the people building the floor are not slowing down.

04 Abundance moves the bottleneck

Here is the twist that the Sol, Terra, and Luna naming makes concrete. As the cost of a unit of intelligence falls toward the floor, intelligence stops being the scarce resource. Judgment about where to apply it becomes the scarce resource instead. The bottleneck moves from the model to the person, or the system, deciding which model to call.

A tiered product line is the supply side admitting exactly that. OpenAI has handed every customer a dial that runs from a dollar to thirty, and the value those customers extract will depend far less on the dial existing and far more on whether they know where to set it for each task. Two companies can buy identical access to the same three models and get wildly different returns, separated entirely by how intelligently they route the work. That gap is the new competitive edge. It is not access. It is allocation.

Access to intelligence is becoming a commodity. Knowing where to spend it is becoming a craft.

05 What a team should actually do about it

The practical shape of this is not complicated, though it does take discipline. Stop defaulting to the biggest model. Sort your work by how hard it really is, and route the easy majority to the cheap tier and the genuinely hard minority to the expensive one. Spend your premium capability where the cost of being wrong is high, which usually means planning, architecture, and the small set of decisions everything else depends on, and economize ruthlessly on the high-volume execution that surrounds them.

Then measure the thing that matters, which is whether the work actually met the bar, not whether it felt fast or felt impressive. And revisit the whole arrangement often, because the prices are not standing still. A routing policy that was optimal last quarter can be wasteful this quarter, simply because the floor dropped again. The teams that win the next few years will be the ones who treat allocation as a living system they tune, not a setting they choose once.

OpenAI named its tiers after the sun, the earth, and the moon, a small solar system you can rent by the token. It is a fitting image for where the field is going. The sun is not better than the moon. It is just for different work, and sending it everywhere would scorch your budget for no extra benefit. The companies that understand this will not be the ones with the most powerful model. They will be the ones who always seemed to know exactly which one to reach for. At superpenguin, that has been the whole point from the start. The only real limit is your budget. Everything interesting happens in how you spend it.