OpenAI on Thursday released the newest version of the artificial intelligence model that powers its popular ChatGPT chatbot, with CEO Sam Altman promoting it as like having a “team of Ph.D. level experts in your pocket.”
The company said the new model, GPT-5, is its smartest and fastest to date with wide-ranging improvements to ChatGPT’s skills in areas like coding, writing and taking on complex actions.
The model has been eagerly anticipated by technologists as a leap forward in AI capabilities and a step toward the kind of superintelligent systems that almost every major technology company is now racing to build. OpenAI along with Microsoft, Meta, Google, Amazon and others have already plowed hundreds of billions of dollars into AI investment, with some projections of future spending reaching into the trillions.
But GPT-5’s new abilities appear to be largely refinements to features that ChatGPT and other AI systems already have. A spokesperson told NBC News that there are still “key limitations in areas like persistent memory, autonomy, and adaptability across tasks.”
GPT-5, which was unveiled to the public during a live streamed event on Thursday, is free for everyone. But users in paid tiers will get higher usage limits, as well as access to GPT-5 Pro.
During a media briefing prior to launch on Wednesday, Altman said he believes the new model is so smart that going back to its predecessor GPT-4, which launched in 2023, feels “quite miserable” in comparison.
“It reminds me of when the iPhone went from those giant-pixel old ones to the retina display, and then I went back to using one of those big pixelated things,” Altman said. “And I was like, ‘Wow, I can’t believe how bad we had it.’”
Starting Thursday, ChatGPT users will also be able to choose a custom color for chats, pick from a selection of pre-set personalities and connect their ChatGPT to Gmail, Google Calendar and Google Contacts.
GPT-5 comes as OpenAI looks to find an edge in the global AI race that has pushed tech companies to roll out sweeping advancements in recent years, with full support from their respective governments. As generative and “agentic” (AI that can autonomously take actions on a user’s behalf) AI tools grow ever more sophisticated, they’ve been met with a mix of excitement and ethical concerns across industries.
The world has seen rapid technological advancements since ChatGPT launched to the public in November 2022, when it was powered by GPT-3.5, and kicked off the generative AI boom that now dominates Silicon Valley.
ChatGPT boasts over 700 million users. Today, powerful large language models are integrated into everything from regular chatbots to agentic assistants and photorealistic video generators.
In Thursday’s stream, OpenAI shared that GPT-5 set a new high on the SWE-Bench Verified, a benchmark that ranks models on their software engineering capabilities.
OpenAI’s post-training lead, Yann Dubois, said the model opens up “a whole new world of vibe coding,” or the ability for noncoders to create fully deployable websites and apps just by describing what they want to build and letting an AI agent do the work.
Altman said he believes GPT-5’s ability to “instantaneously create an entire piece of computer software for you” will likely be a core part of its appeal and use.
“I think this idea of software on demand will be a defining part of the new GPT-5 era,” he said.
Earlier this summer, Altman compared contemporary AI agents to an “intern that can work for a couple of hours.” He went on to predict that “at some point, it’ll be like an experienced software engineer that can work for days.” Such tools have already begun to displace white-collar jobs, rousing distress from job seekers.
GPT-5 dropped amid an already buzzy week for AI companies. OpenAI had just released two new open-weight models, or models that other developers can download and use for free, on Tuesday. Also on Tuesday, Anthropic dropped Claude Opus 4.1, an upgrade to its powerful Opus 4 model, and Google released its interactive world simulation model Genie 3.
Nick Turley, OpenAI’s head of ChatGPT, said the company is confident GPT-5 will be more accurate and produce fewer hallucinations, a term for false or misleading information that AI models present as fact. He said the company has made significant improvements on toning down sycophancy, a problem that has plagued ChatGPT in the past.
The AI giant recently rolled back an update to GPT-4o that made the bot so overly agreeable that it praised users to the point of sounding disingenuous, and even went so far as to endorse and give instructions for terrorism. Just this week, OpenAI also added guardrails to make its models refrain from giving direct advice about personal challenges.
GPT-5 went through 5,000 hours of red teaming, or expert-led safety testing, according to the company’s safety research lead, Alex Beutel. Beutel added that the model will be less deceptive than previous versions, meaning it should be less likely to lie to users about things like completing a task it hadn’t actually completed.
OpenAI also developed a new method, which it dubbed “safe completions,” to prevent people from trying to get around safety guardrails by tailoring their questions in ways that trick the model into aiding in a harmful request.
“So if someone [asks] 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, or it could be a student asking a science question to understand the physics of this material,” Beutel said. “And this creates a real challenge. What’s the best way for the model to reply?”
In the past, ChatGPT would either respond or refuse to respond based on whether it thinks the question is safe to answer. GPT-5, however, will take a new approach: Rather than choosing between compliance and refusal, it will try to always “give as helpful an answer as possible, but within the constraints of remaining safe.”
Online, there’s been some buzz around what GPT-5 will reveal about humanity’s ability to achieve “artificial general intelligence,” or AGI, a hypothetical benchmark at which point AI will be fully capable of doing any intellectual task that a human can.
Altman, who has been vocal about his optimism for the potential to reach AGI, said GPT-5 isn’t there just yet. While the model is generally highly intelligent, he said, it’s still missing a key ability to learn continuously without being retrained.
“But the level of intelligence here, the level of capability, it feels like a huge improvement,” Altman said. “And certainly, if I could go back five years before GPT-3 and tell me we have this now, I’d be like, ‘That’s a significant fraction of the way to something very AGI-like.’”
