OpenAI has released GPT-5, a novel flagship AI model to be used by the company's upcoming generation of ChatGPT.
GPT-5, released on Thursday, is OpenAI's first "unified" AI model and merges the reasoning capacity of its o-series of models with the quick responses of its GPT series. The next-generation model ushers in a new era for ChatGPT — and its developer, OpenAI — and indicates OpenAI's larger goals to create AI systems that are more akin to agents than chatbots.
While GPT-4 has facilitated AI chatbots in providing intelligent responses to a diverse range of inquiries, GPT-5 enhances ChatGPT's capabilities to execute a broad spectrum of tasks on behalf of users, including the generation of software applications, the management of a user’s calendar, and the creation of research briefs.
With GPT-5, OpenAI has also tried to make ChatGPT easier to use. Rather than making users select the correct settings, GPT-5 has a real-time router that determines how to provide the best response, whether that's answering user queries in a speedy manner or taking more time to "think" through responses.
(Image Credits: Zac Wolff / Unsplash)
In a briefing with journalists, OpenAI CEO Sam Altman referred to GPT-5 as "the best model in the world" and stated it was a "meaningful step" on the company's journey to create AI that can beat humans at the majority of economically useful work — in other words, artificial general intelligence (AGI).
Having something like GPT-5 would be almost unimaginable at any prior point in history," Altman said.
Starting Thursday, GPT-5 will become available to all free ChatGPT users as their default model. OpenAI's Vice President of ChatGPT, Nick Turley, said this effort is part of the company's attempt to offer free users access to an AI reasoning model for the first time. (The company previously kept these more advanced models behind a paywall.)
"This is one of the ways that I'm looking forward to living the mission, ensuring that this material actually gets used to help people," Turley said of the choice, alluding to OpenAI's ongoing mission to share sophisticated AI with as many individuals as possible.
The anticipation for GPT-5 is high, one of OpenAI's most highly anticipated product releases since ChatGPT sent the company's profile soaring in 2022. ChatGPT has since become one of the most used consumer products in the world, reaching a weekly user base of more than 700 million—nearly 10% of the global population, according to the company.
GPT-5 is viewed by many as a bellwether for the overall advancement of artificial intelligence, and how this model is received in Silicon Valley could have significant reverberations for big technology companies, Wall Street, and lawmakers who oversee technology. These constituencies are watching to see if GPT-5 offers a significant leap in AI capabilities, as did its predecessor, GPT-4, which shattered expectations for what software could do.
GPT-5 gives a little advantage over the competitors
OpenAI asserts that GPT-5 is state-of-the-art across a number of areas, narrowly outperforming top AI models from Anthropic, Google DeepMind, and Elon Musk's xAI on major benchmarks. GPT-5 narrowly underperforms frontier AI models in some other domains.
The firm states GPT-5 provides frontier-level performance in the realm of coding; Altman claimed the model is particularly good at generating complete software programs on the fly, in what has become referred to as "vibe coding.
On SWE-bench Verified — an exam of actual coding work drawn from GitHub — GPT-5 achieves a score of 74.9% on its very first try. That puts GPT-5 narrowly ahead of Anthropic's new Claude Opus 4.1 model, which achieved a score of 74.5%, and Google DeepMind's Gemini 2.5 Pro, which gained 59.6%.
On Humanity’s Last Exam — a difficult test measuring AI model performance across math, humanities, and the natural sciences — a version of GPT-5 with extended reasoning (GPT-5 Pro) scored 42% when using tools. That’s slightly less than xAI was able to achieve with Grok 4 Heavy, which scored 44.4% on the test. On GPQA Diamond — a exam of PhD-level science questions — GPT-5 pro achieved 89.4% on the first attempt, beating Claude Opus 4.1 at 80.9% and Grok 4 Heavy at 88.9%.
OpenAI claims GPT-5 is improved at responding to health questions. In a test for accuracy in AI model responses on healthcare subjects, HealthBench Hard Hallucinations, OpenAI claims that GPT-5 (with thinking) hallucinates only 1.6% of the time. That is much lower than the company's earlier GPT-4o and o3 models, which rated 12.9% and 15.8, respectively.
Though AI chatbots are not doctors, millions of individuals are turning to them for medical advice. Responding to this trend, the company reports that GPT-5 is more proactive in raising red flags for the user's health and assisting users in interpreting medical results.
Furthermore, OpenAI claims GPT-5 is superior to other AI models on more challenging-to-quantify, subjective areas, like creative writing and design. Turley claimed GPT-5 generates responses in a more natural way and has "better taste" than other AI models when it comes to creative tasks.
Turley said, "The vibes of this model are really good.
GPT-5 is also more precise than OpenAI's earlier models, and the firm reports it suffers much less from hallucinations — the inclination for AI models to invent information — than its o-series models. Hallucinations appeared to be deteriorating in OpenAI's newest AI reasoning models, like o3, and OpenAI earlier stated it didn't quite comprehend why it was occurring.
For answers to ChatGPT prompts, OpenAI discovered that GPT-5 (with thinking) hallucinates and answers with false information 4.8% of the time. That is a big drop from o3 and GPT-4o, which have hallucination rates of 22% and 20.6%, respectively, on the test.
On a benchmark of an AI model's agentic capability to carry out simulated online tasks, Tau-bench, GPT-5 has mixed performance. On a section of the test for an AI's capability to navigate an airline company's website, GPT-5 performs 63.5%, slightly worse than o3, which achieved 64.8%. On another section of the test for AI's capability to navigate e-commerce websites, GPT-5 performs 81.1%, worse than Claude Opus 4.1, which achieved 82.4%.
OpenAI claims that GPT-5 is more secure than its predecessor models. While AI reasoning models occasionally exhibit a tendency to come up with plots against humans or to lie in order to advance their own agendas, OpenAI found that GPT-5 exhibited deceptive behavior at a lower frequency than other models.
OpenAI safety research lead Alex Beutel said that reducing deception improves not just the safety of GPT-5 but also the user experience overall, thus creating a model that is more "transparent and honest in ways users can trust."
Beutel also adds that GPT-5 is more capable of distinguishing between malicious actors attempting to use ChatGPT for nefarious purposes and users asking innocuous questions. This allows GPT-5 to reject more unsafe queries, while providing fewer rejections for users asking innocuous questions.
Improvements for consumers and developers.
ChatGPT is receiving some user experience enhancements as part of the GPT-5 rollout. Now, users can choose among four new personalities in ChatGPT's setting: Cynic, Robot, Listener, and Nerd. The organization states these will adjust ChatGPT's responses without the need for users to explicitly ask the model to answer in a particular manner.
Plus subscribers to ChatGPT, which costs $20 a month, have higher usage limits on GPT-5 than non-paying users. Meanwhile, those subscribed to the $200-a-month Pro tier will get unlimited use of GPT-5, along with a more refined version called GPT-5 Pro that uses more computational resources to produce better answers. Companies signed up for OpenAI's Team, Edu, and Enterprise plans will get access to GPT-5 as their default model within the next week.
For developers, GPT-5 is arriving in OpenAI's API in three sizes — gpt-5, gpt-5-mini, and gpt-5-nano — that will take more or less time "reasoning" through tasks. Developers can now also control verbosity in the OpenAI API, determining how long or short an AI model's responses are.
The GPT-5 base model will charge developers $1.25 for every million input tokens (about 750,000 words, longer than the whole "Lord of the Rings" trilogy) and $10 for every million output tokens. The release of GPT-5 comes after a week of high activity for OpenAI. The company announced an open-weight reasoning model, gpt-oss, that developers and businesses can download for free and run at a greatly diminished cost.
That open model came very near to the performance of OpenAI's previous top models, o3 and o4-mini, but GPT-5 sets a new standard for state-of-the-art performance in some areas, especially coding. Nevertheless, GPT-5 appears to be approximately equivalent to other leading AI models across various domains. It is important to note that benchmarks provide only a partial understanding of any AI model's capabilities. Consequently, it remains uncertain how developers will implement GPT-5 in practical applications, as well as whether this model constitutes a genuine advancement over its competitors.
Adapted from reporting by TechCrunch, with added commentary and analysis by SNDmix.