ChatGPT-5 has been trained to minimise hallucinations and avoid deception.

IMAGE: Sam Altman, CEO, OpenAI. Photograph: Amir Cohen/Reuters

Sam Altman-led OpenAI on Thursday launched its most advanced large language model (LLM), ChatGPT-5, making it freely available to users worldwide through both its chatbot interface and application programming interface (API).

Among the large language model’s (LLM’s) key features is its ability to generate working software instantly: Users can describe, in a series of prompts, the web application they want to build, and the model will write the code for them.

‘This is a major upgrade over 4.0 and a pretty significant step along our path towards AGI (artificial general intelligence),’ said OpenAI Chief Executive Sam Altman.

It’s the first time one of the company’s mainline models feels like a ‘legitimate expert’, he said.

The San Francisco-based firm had launched its first model, ChatGPT 3.5, back in November 2023.

Since then, the company’s LLMs have evolved rapidly and helped OpenAI become one of the most influential players in the global tech ecosystem.

Its LLMs have more than 700 million weekly users globally.

Nick Turley, head of ChatGPT at OpenAI, said the new model has lower rates of hallucination than even the o3 model.

‘If you have been using o3, you might notice that while it is very intelligent, it can go off thinking for quite some time.

‘And GPT 5 can also think, but it will do so much more quickly, so you have the best of both worlds.

‘You have it reason when it needs to reason, but you do not have to wait as much longer,’ Turley said.

Among the model’s most striking abilities is what Altman described as ‘good instantaneous software’.

The tool also excels at designing artificial intelligence-enabled agents, he said.

Turley elaborated on this functionality, saying users can now walk ChatGPT-5 through a step-by-step description of a web application they want to build, and the model will write the relevant code in response.

For example, a user might wish to create an app to help teach French to non-French speakers.

By prompting the model with the desired features, ChatGPT-5 will generate executable code within the same window.

‘Even though this seems pretty simple, building such an interactive website is actually pretty complex and would require multiple hours for a software developer,’ said Yann Dubois, a member of OpenAI’s technical staff, during a demonstration.

ChatGPT 5, Dubois said, would open ‘a new world of VibeCoding’, where users would not be required to be experts at writing code or knowing programming languages to create web applications with ease.

The model has been designed with a strong emphasis on safety, according to Alex Beutel, who leads OpenAI’s Safety Research team.

ChatGPT-5, he said, has been trained to minimise hallucinations and avoid deception — instances where a model might falsely claim to have seen an image or completed a task.

‘We have built evaluations to try and track the prevalence of this kind of behaviour and trained the model to do it less,’ Beutel said.

‘We see much lower rates of deception in GPT-5 compared to past models.’

To further improve user safety, especially in ambiguous situations where it is unclear whether a prompt is aimed at acquiring knowledge or causing harm, the model will adopt a method called ‘safe completions’.

This approach involves giving clear but limited answers that remain within boundaries deemed secure.

‘The model might only partially answer (such questions), sticking to a higher level of detail that can’t actually be used to cause harm.

‘We think this is a better foundation for safety generally, especially in these dual-use cases,’ Beutel explained.

The new model, Altman said, will be far beyond the frontier of any other model out there, and help make the world more accessible.

‘There was another mission as a company, which says that we are willing to forego profits to make AGI broadly beneficial to everyone.

‘We are willing to do these things that other companies might not,’ he added.

Feature Presentation: Ashish Narsale/Rediff