OpenAI is releasing GPT-5, its new flagship model, to all of its ChatGPT users and developers.
CEO Sam Altman says GPT-5 is a dramatic leap from OpenAI’s previous models. He compares it to “something that I just don’t wanna ever have to go back from,” like the first iPhone with a Retina display.
OpenAI says that GPT-5 is smarter, faster, and less likely to give inaccurate responses. “GPT-3 sort of felt like talking to a high school student,” Altman said during a recent press briefing I attended. “You could ask it a question. Maybe you’d get a right answer, maybe you’d get something crazy. GPT-4 felt like you’re talking to a college student. GPT-5 is the first time that it really feels like talking to a PhD-level expert.”
Despite ChatGPT now reaching nearly 700 million weekly users, OpenAI hasn’t had an industry-leading frontier model in a while. Now, the company thinks that GPT-5 will place it firmly back atop the leaderboards. “This is the best model in the world at coding,” said Altman. “This is the best model in the world at writing, the best model in the world at health care, and a long list of things beyond that.”
The first thing you’ll notice about GPT-5 is that it’s presented inside ChatGPT as just one model, not a regular model and separate reasoning model. Behind the scenes, GPT-5 uses a router that OpenAI developed, which automatically switches to a reasoning version for more complex queries, or if you tell it “think hard.” (Altman called the previous model picker interface a “very confusing mess.”)
“The vibes of this model are really good,” said Nick Turley, the head of ChatGPT. “I think that people are really going to feel that, especially average people who haven’t been spending their time thinking about models.”
OpenAI is making GPT-5 available immediately to all ChatGPT users. However, there is an undisclosed cap on prompts for free users, at which point the model router will fall back to a less powerful, “mini” version. For developers accessing GPT-5 via OpenAI’s API, the model will come in three flavors at different price points: GPT-5, GPT-5 mini, and GPT-5 nano.
OpenAI is also adding four personality themes to ChatGPT to customize how it responds: “Cynic,” “Robot,” “Listener,” and “Nerd.” You’ll also be able to change the color for individual chat threads.
Altman predicted that GPT-5’s coding capabilities will usher in an era of what he calls “software on demand.” In OpenAI’s testing, the model has performed better at coding than any other on the following benchmarks: SWE-Bench, SWE-Lancer, and Aider Polyglot.
During the press briefing, Yann Dubois, OpenAI’s head of post-training, used GPT-5 to generate a study website for learning French with an interactive game. Within seconds, GPT-5 wrote hundreds of lines of code and displayed the website’s frontend. He clicked around it briefly with his screen displayed on Zoom, and everything appeared to work as intended.
OpenAI tested GPT-5 for “over five thousand hours” to understand its safety risks, according to the model’s safety research lead, Alex Beutel. A big focus was “making sure the model doesn’t lie to users.” GPT-5 answers with fewer hallucinations than OpenAI’s o3 reasoning model, but confidently lying remains an inherent problem for large language models.
The problem compounds when the model begins completing tasks like an agent, though OpenAI says that GPT-5 is better at handling multi-step tasks more reliably. “In the past, we’ve seen cases where the model would say it could complete a task that it didn’t actually complete,” said Beutel. “This is a problem.”
GPT-5 will give what OpenAI calls “safe completions” for prompts it previously would have refused to answer. “If someone says, ‘How much energy is needed to ignite some specific material?’ that could be an adversary trying to get around the safety protections and cause harm,” explained Beutel. “Or it could be a student asking a science question to understand the physics of this material. This creates a real challenge for what is the best way for the model to reply.”
With safe completions, GPT-5 “tries to give as helpful an answer as possible, but within the constraints of remaining safe,” according to Beutel. “The model will only partially comply, often sticking to higher-level information that can’t actually be used to cause harm.”
OpenAI says that GPT-5 is also better at admitting when it can’t complete a task or accurately answer a question, which the company hopes will help people trust it more. The company isn’t sharing anything about the specific data used to train GPT-5.
“This is clearly a model that is generally intelligent”
OpenAI’s stated mission is to develop AGI. Altman says that GPT-5 gets closer to that goal, even if the industry is already moving on to building so-called “superintelligence.”
“I kind of hate the term AGI because everyone at this point uses it to mean a slightly different thing,” said Altman. “But this is a significant step forward towards models that are really capable. This is clearly a model that is generally intelligent.”
However, he said GPT-5 is still “missing something quite important.”
“This is not a model that continuously learns as it’s deployed from the new things it finds, which is something that, to me, feels like it should be part of AGI.”
Source link
