Yesterday OpenAI unveiled GPT-5.6, its new family of models: Sol, the powerful one, Terra, the balanced one, and Luna, the cheap one. And the numbers point to something big. Sol leads on several heavyweight tests: long-horizon agentic tasks, tool use, web browsing. On ARC-AGI-3, the benchmark where almost every model crashes, it was the first frontier model to solve one of its puzzles.
It is, without a doubt, one of the most capable AIs out there today.
Now, a detail OpenAI did not highlight: on the general Artificial Analysis index, the independent reference, Anthropic's Fable 5 is still ahead by a little. It is not even the clean number 1. And almost every other number in the announcement comes from OpenAI itself.
It doesn't matter. Because even if it were the undisputed number 1, almost nobody would need it.
Two stories that prove the opposite of what they wanted to prove
For the launch, OpenAI published a video with three use cases. A farmer in Japan who built remote control for his greenhouse with a Raspberry Pi. A couple in New York who launched a cereal-box business. And a mathematician in Poland who had been stuck on a conjecture for 3 years and finally unlocked a new idea.
Look closely at all three.
The first two are tasks that have been solved for months with far cheaper models. Neither needs the best model in the world. Only the third, the mathematician, truly justifies that much power. As he put it himself: if you have the audacity to attempt something huge, you stop being afraid of compute because you know how to organize it.
That is the point. The vast majority of us are not solving impossible conjectures. We use the tools in everyday ways. And that is fine.
The question changed, and almost nobody noticed
For two years the question was "which AI is the smartest?". It was a collector's question. The answer changed every few months and was never much use.
The good question is another one: which model is good enough for this specific task at the lowest cost?
That is why OpenAI did not launch one model, it launched three. Sol costs $30 per million output tokens. Luna, $6. Five times less. And to write an email, sort out some ideas or generate a first draft of something, Luna is more than enough. Paying for Sol there is like buying a 15,000-euro Hasselblad for your holiday photos.
Nobody gives you a prize for using the most expensive model. They give it to you for delivering the work well and on time.
In visual production, exactly the same thing happens
This is not about text AI. It is about judgment.
When we produce an image or a video for a brand, we have in front of us a catalog of engines that rotates every week. One nails skin. Another nails the product but invents the hands. Another is wildly expensive and only pays off in the campaign's hero shot. The client sees none of that, nor should they. The client sees the result.
Our job is not to have access to the top model. Anyone with a credit card has that. Our job is to know which one is enough for each piece, and not to spend a cent more on the ones that do not ask for it.
Whoever collects models shows off the tool. Whoever produces with them delivers the result. They are not the same thing, and it shows in the invoice.
And tomorrow, this will be the norm
The best part of all this is that it expires fast. Today's frontier model is tomorrow's everyday model. Two years from now you might use GPT-5.6 Sol to fix an email without a second thought, just as today you use for that models that three years ago were science fiction.
So no, GPT-5.6 is not a story about who has the smartest AI. It is a reminder that the edge was never in the tool. It is in knowing which one to use, when, and for how much.
The most expensive is rarely the smartest. And the smartest is almost never what you need.
Source
- OpenAI, GPT-5.6: Frontier intelligence that scales with your ambition (July 9, 2026): https://openai.com/index/gpt-5-6/