Openai’s GPT-5 is here | TechCrunch

Openai has launched the GPT-5, the new flagship AI model that bolsters the company’s next-generation ChatGPT.

Released Thursday, the GPT-5 is OpenAI’s first “unified” AI model, combining the model’s O-series inference capabilities with the fast response of the GPT-series. The next-generation model marks a new era of ChatGpt and its creator, Openai. It points to Openai’s broader ambitions to develop AI systems that resemble agents rather than chatbots.

GPT-4 allows AI chatbots to provide smart responses to a variety of questions, while GPT-5 allows ChatGPT to complete a variety of tasks on behalf of the user, such as generating software applications, navigation of the user’s calendar, and creating research briefs.

With GPT-5, Openai also wants to make using ChatGpt easier. Instead of asking users to choose the right settings, the GPT-5 is equipped with a real-time router that determines how to provide the best answer, whether it promptly responds to a user’s questions or spends extra time “thinking” through the answer.

During a briefing with reporters, Openai CEO Sam Altman claimed that the GPT-5 is the “best model in the world” and described it as a “critical step” along the company’s path to developing AI that can outperform humans in the most economically valuable work: artificial general information (AGI).

“It’s hardly ever imagined in history to have something like the GPT-5,” says Altman.

Starting Thursday, the GPT-5 will be available as the default model for all free users of CHATGPT. Nick Turley, Vice President of Openai’s ChatGpt, said this is part of the company’s efforts to make AI inference models accessible to free users for the first time. (Previously, the company had a grasp of these more advanced models behind the paywall.)

“This is just one of the ways I’m excited to live my mission. Make sure that something like this actually benefits people,” Turley talks about the decision, referring to Openai’s long-standing mission of distributing as many people as possible.

The GPT-5 is highly anticipated, and it has been one of Openai’s most predicted product launches since ChatGPT put the company on the map in 2022. Since then, ChatGPT has grown into one of the world’s most popular consumer products, reaching over 700 million users each week.

Many view the GPT-5 as AI positive, and Silicon Valley’s model reception could have a major impact on policymakers regulating technology. These stakeholders are seeing whether the GPT-5, like its predecessor GPT-4, offers a big jump to AI’s capabilities, as it challenged the expectations that the software could do.

The GPT-5 offers a slight edge in the competition

Openai is cutting edge in several domains, with a slight border on Xai’s major AI models from humanity, Google Deepmind and Elon Musk in key benchmarks. However, the GPT-5 is slightly below the frontier AI models in other regions.

According to the company, the GPT-5 offers frontier-level performance in terms of coding. Altman said the model, what became known as “Vibe Coding,” is particularly good at spinning up entire software applications on demand.

Testing the actual coding tasks drawn from GitHub in SWE-Bench validation – GPT-5 scored 74.9% on the first attempt. So the GPT-5 simply outperforms Anthropic’s latest Claude Opus 4.1 model, which won 74.5%, and Google Deepmind’s Gemini 2.5 Pro, which won 59.6%.

The final exam of humanity – a difficult test to measure the performance of AI models in mathematics, humanities, and natural sciences – the version of GPT-5 with extended inference (GPT-5 Pro) won 42% when using the tool. This was slightly less than Xai was able to achieve with the Grok 4 Heavy, earning 44.4% in Tests.

In GPQA Diamond, a test of PHD-level science questions, the GPT-5 Pro scored 89.4% in the first attempt, surpassing Claude Opus 4.1, earning 80.9%, and Grok 4 Heavy scored 88.9%.

Openai says GPT-5 is good for answering health-related questions. Healthcare topic, Healthbench Hard Hallucinations, Openai, GPT-5 (with thought) hallucinations are only 1.6% of the time, and in a test to measure the accuracy of the response of AI models. This was much lower than the company’s previous GPT-4O and O3 models, earning 12.9% and 15.8, respectively.

AI chatbots are not medical professionals, but millions of people use them for their health advice. In response to this phenomenon, the company says that GPT-5 is more proactive about flagging potential health concerns and helping users analyze medical results.

Furthermore, Openai says that the GPT-5 is superior to other AI models in more difficult subjective domains, such as creative design and lighting. According to Turley, the GPT-5 responds naturally and shows a “better taste” than other AI models for creative tasks.

“The vibe of this model is really good,” Tarly said.

The GPT-5 is more accurate than Openai’s previous models, and the company says that compared to the O-series model, AI models are far less likely to form information. Hallucinations appeared to be exacerbated with Openai’s latest AI inference models such as O3.

In response to the ChatGpt prompt, Openai found that GPT-5 (including thoughts) hallucinations responded in 4.8% with false information. This is a significant decrease from O3 and GPT-4O, with hallucination rates of 22% and 20.6% in tests, respectively.

In benchmarks that measure agent capabilities in AI models, the TAU bench, GPT-5, provides mixed performance to complete simulated online tasks. In some tests measuring AI’s ability to navigate airline websites, the GPT-5 achieved 63.5% and slightly reduced O3, which achieved 64.8%. In another part of the test that measures AI’s ability to navigate retail websites, the GPT-5 scored 81.1%, beating the Claude Opus 4.1 and 82.4%.

Openai also says that the GPT-5 is safer than its previous models. AI reasoning models sometimes show tendencies to scheme towards humans or lie to promote their own goals, but Openai has found that GPT-5 is deceptive at a lower rate than other models.

Openai’s Safety Research Leader Alex Beutel said that reducing deceptions will not only improve the safety of the GPT-5, but also the user experience, creating a model that is “transparent and honest in a way that users can trust.”

Beutel also notes that GPT-5 is excellent at identifying between bad actors trying to misuse ChatGPT and users making harmless requests. This allows GPT-5 to reject unsafe questions and does not provide rejections to users seeking harmless information.

Upgrades for Consumers and Developers

CHATGPT has obtained several user experience upgrades as part of launching GPT-5. Users can now choose from four new personalities (sarcasm, robot, listener, nerd) in the CHATGPT settings. The company says these adapt ChatGpt responses without requiring users to specifically ask the model to respond in a specific way.

Subscribers to the ChatGpt $20/month plan have a higher GPT-5 usage limit than free users. Meanwhile, Pro subscribers, who cost $200 per month, have unlimited access to GPT-5 and a souped-up version called GPT-5 Pro, which uses additional computational resources to generate better answers. Openai’s teams, EDUs, and Enterprise Plan organizations will have access to GPT-5 as their default model for next week.

For developers, GPT-5 appears in OpenAI’s API in three sizes: GPT-5, GPT-5-MINI, and GPT-5-NANO. Additionally, developers can control redundancy with the Openai API and determine how long or short the response of the AI model will be.

The GPT-5 base model costs developers $1.25 input tokens per million (approximately 750,000 words, the entire “Rod of the Rings” series) and $10 output tokens per million.

The release of the GPT-5 comes after Openai’s busy week. The company has released the open weight inference model GPT-Oss, which allows developers and companies to download for free and run for a fraction of the cost. The open model roughly coincided with the capabilities of Openai’s previous top models, the O3 and O4-Mini, but the GPT-5 sets new standards for frontier performance in some regions, such as coding.

However, the GPT-5 appears to be roughly comparable to other frontier AI models in several areas. Of course, the benchmark only tells a part of the AI model’s story, and it remains to be seen how developers will use GPT-5 in the real world, and whether the model is truly a competitive step.