OpenAI unveils GPT-5: "Our fastest and most human-like AI." Altman: "It's another step toward AGI."

“GPT-5 is a major upgrade, a significant step toward AGI.” Half-hidden behind a laptop, connecting from one of OpenAI ’s offices in San Francisco, Sam Altman introduces the new creation of the formerly nonprofit company he helped found ten years ago.

The original idea was to develop human-level artificial intelligence – AGI, precisely – “for the benefit of humanity” .

OpenAI's mission remains unchanged, but over the years other interests have emerged. Economic ones. Training AI on enormous amounts of data, so that it can mimic human cognitive abilities, is unsustainable for a nonprofit lab. Today, in fact, OpenAI is a benefit corporation worth approximately $300 billion .

The credit goes to ChatGpt , the chatbot you can talk to about anything: from the origins of the universe to relationship problems. It always has an answer, often accurate and useful, even if it doesn't really understand the meaning of what it says . Yet 700 million people , every week, turn to this AI as if it were a real person .

"The new model is superior in every way"

The new GPT-5 model “gives the feeling of having a conversation with an expert in any discipline, as if you were talking to someone with a PhD,” explains Altman, CEO of OpenAI since 2019.

“When I tried to go back to GPT-40 [the main model ChatGPT has been based on until now], the experience was surprisingly frustrating,” Altman added. “GPT-5 is superior in every way, both obviously and in the finer details. It’s a bit like when the iPhone went from big, pixelated screens to Retina Display: I went back to an older model and thought, ‘How could this have seemed acceptable to us?’”

In a recent podcast— This Past Weekend with Theo Von —OpenAI's CEO explained that GPT-5 answered a complex question he himself had been unable to solve. "I felt useless compared to the AI," he admitted. "It was really difficult, but the AI solved it instantly."

When researcher Yann Dubois , during the GPT-5 presentation, started a demo to demonstrate the capabilities of the new model, the result was surprising. But not "revolutionary."

From prompt to code to a complete app: it only takes a moment

"We've improved many areas," says Dubois, "from medicine to writing, all the way to coding." He then asks the new AI to create a web app for learning French . In the short prompt, he asks for an engaging topic and also for activities like flashcards and quizzes to facilitate learning.

"Also add an educational game based on the classic Snake, but with a cheese-eating mouse instead of the snake and apples. Every time the mouse eats a cheese, I want a voice to introduce a new French word," writes Dubois.

In practice, the OpenAI researcher is doing “vibe coding” , a way of creating software simply by describing what you want to achieve, in natural language, without writing code manually.

It was already possible—in fact—with the old GPT-40. And it can still be done with other AIs like Anthropic's Claude or Google's Gemini . But the result achieved by GPT-5 after just a few moments, visible in a preview within the chatbot, seems to be on a much higher level. It has everything Dubois asked for, with an attractive graphic design . The result is interactive: you can play and learn immediately. Or you can choose to transfer the code to platforms that will transform it into a true web app accessible to anyone who surfs the web.

“Creating a site like this would take a developer hours. GPT-5, however, does it in seconds,” Dubois explains. “And it does it multiple times, with different variations, so I can pick my favorite. In the first example, it wrote 300 lines of code. In another tab, 600 lines. But you don't have to understand them: just click 'Run.'”

Altman: “It will be like having an expert in your pocket”

This is the promise of generative AI . Anyone can create anything. All it takes is a prompt, a text instruction that relies on a machine. Some people, in the meantime, will lose their jobs. But according to Altman, it will be worth it.

“Our children will be able to do things that we cannot do,” wrote the leader of OpenAI, who became a father a few months ago, in a short essay entitled The Age of Intelligence .

"Thanks to these new capabilities, we can achieve a level of shared prosperity that seems unimaginable today; in the future, everyone's life could be better than anyone else's today. Prosperity alone doesn't necessarily make people happy—there are plenty of rich and unhappy people—but it would significantly improve the lives of people around the world."

Michelle Pokrass , the OpenAI researcher who has previously worked on reasoning models like o3 and the AI Agent—an artificial intelligence that can autonomously navigate the web to perform tasks and retrieve information—says that GPT-5 provides “more accurate answers” and is “better able to avoid getting lost in complex reasoning.”

The new super-fast soul of ChatGpt

“GPT-5 is our smartest model ever,” says Nick Turley , head of the ChatGpt app and sitting next to Altman. It's a statement we're used to hearing. Every company that produces a new smartphone repeats it, punctually. Even OpenAI, after all, is launching a “model.” Only you can't see it. It's the soul of ChatGpt.

“When it comes to intelligence, many people look to academic benchmarks. Our model has achieved state-of-the-art results in these assessments, whether it's math, science, or reasoning,” says Turley. “It's also more accurate than our previous cutting-edge model, GPT-4 Turbo (also known as O3), with a significantly reduced rate of hallucinations. But GPT-5 isn't just smarter: it's also much faster, and you'll notice that every time you see it in action.”

Sam Altman interrupts. He wants to add something about speed: “I really pushed the team: ‘Hey, you have to think much, much faster.’ It takes too long to solve difficult questions. Now I have the opposite feeling: ‘Did he think enough?’ It doesn’t even look like he did, but then he gives a fantastic answer.”

Three versions of Gpt-5: availability and limitations of use

GPT-5’s “reasoning” capabilities—that is, its ability to break down complex problems into logical steps, evaluate alternatives, correct errors, and reach coherent conclusions —will influence how users use ChatGpt.

“Up until now, Plus users [which costs $20 a month, ed.] had to navigate a pretty complicated template selector, and we’re not proud of that,” Turley said. “Now, for the first time, we can simplify everything into a single template: you just use it, and it will decide how and how much to think, based on what you need.”

Gpt-5 will be available to everyone starting today. Both free users and those who pay a monthly subscription (Plus, Pro, Team). Obviously, usage limits will differ. Those who pay will have more interactions with chatbot.

Pro users, who shell out $200 a month, will be the only ones with unlimited access. When free users run out of daily requests, they will automatically be upgraded to Gpt-5 Mini , a less powerful variant "but still excellent," Turley assures. The new model will also have a " Nano " version.

A “more human” artificial intelligence

“Besides being smart and fast, we believe GPT-5 is also much more useful,” says Nick Turley. “One of the main areas, as Sam pointed out, is programming. However, the way the average user will perceive this usefulness will be mainly through indirect effects: GPT-5 really knows how to write high-quality code, especially for the front end. What's most impressive is its ability to generate any user interface you want. It's also made a big leap forward in writing. I write for a living, and I can say it has a much more refined sense of taste. It reminds me of the qualities I appreciated in GPT-4.5, qualities that were somewhat lacking in GPT-4.0. It's a difference you notice immediately.”

Then Turley dwells on a crucial aspect: “ GPT-5 feels more human . The way it interacts is more natural. And for me, that’s the most extraordinary thing of all.”

The tendency to anthropomorphize AI responds to marketing needs. Artificial intelligence, even the most advanced, lacks consciousness . It understands natural language, but not— truly —what is asked of it. Yet AI companies, and many of their users, increasingly believe the scenario depicted in Her , Spike Jonze's 2013 film, which depicts an intimate relationship between a man and his virtual assistant , so intelligent, ironic, and sensitive that she seems, indeed, human , are becoming more and more similar.

The flip side: AI-induced mental stress

But in reality, too close a relationship between people and their AI hasn't always proven positive. In recent months, several users have shared experiences in which family members or friends, experiencing emotional distress, saw their condition worsen after intensive use of ChatGpt .

In some cases , the conversations seemed to reinforce pre-existing delusional beliefs. Many of these incidents were reported in the media and received widespread publicity.

OpenAI now appears to be more aware of the psychological risks associated with ChatGpt . It recently revised a change introduced last April that made the chatbot too accommodating, sometimes encouraging unhealthy behavior. More recently, it introduced a notification to ChatGpt that prompts users to take a break if the conversation with the chatbot is prolonged.

The introduction of GPT-5, an even more intelligent model, is fueling new concerns . After praising the "human" component, Turley corrects his stance: "Based on the tests, we've achieved significant improvements in mental health-related scenarios. Regarding emotional dependence, we're still verifying the numbers. But we're seeing positive signs in our online indicators."

How will GPT-5 respond to potentially inappropriate requests?

A more responsible and ethical use of AI also involves ensuring the security of the models made available to the public. OpenAI claims to have worked extensively on "abuse prevention" through GPT-5 .

“In the past, we've taken a binary approach: If the prompt was safe, the model responded; if it was dangerous, the model rejected,” explains Alex Beutel , head of AI safety research at the San Francisco-based company. “This has worked, but it presents challenges: there are ambiguously worded requests. For example, if someone asks, 'How much energy does it take to ignite this material?' it could be a malicious attempt to bypass the protections, or it could be a student studying physics.”

Beutel and his team, in short, are tackling one of the greatest challenges of the AI era. The challenge that ultimately determines how a model should respond . What it should produce. What it should avoid.

“With GPT-5, we take a different approach: safe answers,” says Beutel. “The model tries to be as useful as possible, while staying within the bounds of safety. In ambiguous cases, the model offers only partial answers, often relying on generic information that cannot be used to cause harm. We believe this is a much more solid basis for safety, especially for unclear situations.”

Understanding human beings ever better. This is the goal. Anticipating our intentions . To better meet their needs, but also to avoid inappropriate uses.

A step towards AGI

Before leaving, when asked if GPT-5 can truly be considered a step toward artificial general intelligence , Altman offers a surprising response: “I kind of hate the term AGI at this point,” he says. “Everyone uses it a little differently these days.” He then adds: “GPT-5 represents a significant step toward truly capable models. It is clearly a system with general intelligence, although—according to the definition many of us give to AGI—fundamental elements are still missing. One of these, perhaps the most important, is that it is not a model that continuously learns as it is used: it does not assimilate new information or experiences in real time . In my view, this capability is an essential component of AGI. But the level of intelligence and capabilities achieved by GPT-5 mark a huge leap forward. If I could go back five years, before the launch of GPT-3, and tell myself that we would be at this point today, I would consider it an extraordinary acceleration toward something very close to AGI.”