GPT-5.6 is probably the best AI model in the world. And precisely for that reason, the majority does not need it.

Yesterday OpenAI publicly released GPT-5.6its new family of AI models with three variants: Sol, the most powerful, Terra, more balanced, and Luna, the most cost-efficient. One idea stood out in that release: that GPT-5.6 is probably the best model in the world.

And precisely for that reason, the vast majority will never need it.

As part of the launch, he published a nice video in which he showed how a farmer in Japan, an entrepreneurial couple in New York and a mathematician in Poland had used it for their work.

Then we go back to the video and those three scenarios. But a preview: two of those three stories demonstrate just the opposite of what OpenAI wanted to demonstrate.

In the official announcement OpenAI also told us about how this was the most capable family of AI models they had ever released and they included the traditional huge string of internal benchmark results to prove it.

Their data revealed that we are facing what theoretically it is the best AI model in the world currently. And the interesting thing is that independent studies like those of Artificial Analysis They corroborate it: in several of its tests GPT-5.6 even surpassed Fable 5, Anthropic’s frontier model that until now was the great reference in this industry.

The model certainly appears to be spectacular. Those responsible for ARC Prize, that benchmark in which most AI models repeatedly crash, commented how GPT-5.6 Sol was still the first frontier model to solve one of the puzzles of their new benchmark, ARC-AGI 3.

No other had come close to that milestone, and according to this organization “it is the best model when it comes to orienting yourself in a situation that you have never encountered.” All that these tests validate is the idea that we are facing a prodigious model.

And the problem is precisely that: that most users will probably never need it.

Too powerful for most of us

Let’s go back to the video at the beginning. Of the three use cases mentioned, two are quite trivial. GPT-5.6 helped the Japanese farmer create a remote control system for his greenhouse with a Raspberry Pi. He helped the New York couple build a curious cereal box business. Nothing in those two tasks seems to require the best model in the world. In fact, they are precisely the type of projects that have been being resolved for months with much cheaper models.

With the third scenario, that of the Polish mathematician, things change: this academic was trying to solve a conjecture that he had been working on for three years. No previous model had been able to help him, but with GPT-5.6 he managed to reveal a totally new idea, he says. One of his final comments precisely makes it clear who GPT-5.6 Sol is for:

“If you have that kind of audacity to try to do something really big, you won’t be scared of the incredible computing power because you can organize it with the model.”

That is the key to the issue: most users are not trying to solve mathematical conjectures that are almost impossible to solve. Most we use tools in a much more everyday wayand that is completely logical and reasonable. That’s why there are many more more modest and affordable models, and why the GPT-5.6 Sol, even if it makes sense, will be a very unprofitable model for most people. It is not a model for counting R’sof course.

In fact, every time a new model comes into our hands, It is very difficult to appreciate if it is really better than the previous ones because the tasks we propose are usually solved very well with the existing ones. There are cases in which differences are seen—in especially complex programming, for example—but here we are faced with a situation that we have lived before on several occasions.

This happens, for example, with modern hardware: very few people need the most advanced processors or an RTX 5090 to play, because more modest CPUs and GPUs give access to a truly fluid experience. We don’t usually need a camera either. Hasselblad of 15,000 euros for our vacations, and a good cell phone of 500-1,000 euros at this point solves the problem wonderfully.

The good thing about all this is what also happens with those examples that we mentioned before: what is extraordinarily expensive and powerful today will end up no longer being so because other even better (and probably more expensive) AI models will appear on the horizon.

The question is no longer “which model is more “intelligent”?”, but “Which model is smart enough to solve this task at the lowest cost?“. That explains why there are variants like Sol, Terra and Luna.

Maybe in two years GPT-5.6 Sol will be the cheap model we use to correct an email or plan a vacation. The recent history of AI invites us to think precisely that: today’s frontier models end up becoming tomorrow’s everyday models.

Perhaps that is the true meaning of GPT-5.6. Not that today almost no one needs so much intelligence, but that in a few years we will probably we all take it for granted.

In Xataka | OpenAI just launched GPT-Live: ChatGPT voice mode has learned to listen, shut up and respond better