OpenAI announces GPT-5 - "dramatically better" and available to all

Today, on 7 August, OpenAI presented GPT-5, the fifth generation of its artificial intelligence. At the presentation, Sam Altman called it "dramatically better" than anything else currently available and said that you won't want to go back from it. He compared the language models as follows: GPT-3 was like talking to a high school student - sometimes it was the right thing to say, but sometimes it was complete nonsense; GPT-4 was like talking to a college student; with GPT-5, you'll be talking to a PhD-level expert. Altman stated that this is the best model for programming, writing, healthcare, and much more.

The new model does not have separate options for in-depth analysis or performance. OpenAI has developed a special analyser that independently determines how simple a question is and, depending on this, selects a "thinking" model. The user can hint the AI to use deep analysis by specifying "think carefully" in the query. For free users, there is a limit after which the model will switch to the simplified GPT-5 mini thinking model.

Paid users who use the API for their own applications or services can choose the model manually. The models available are GPT-5, GPT-5 mini, and GPT-5 nano.

AI now has four personality types: cynic, robot, sensitive, and nerd. You can also change the colour of the window for each chat.

The GPT-5 model was tested by SWE-Bench, SWE-Lancer, and Aider Polyglot benchmarks, where it showed the best results of all existing ones.

A new era in programming has begun. At the presentation, ChatGPT was asked to write a website for learning French in the form of an interactive game - and the model wrote hundreds of lines of code in seconds, which created a website that looked quite working at the presentation.

For more than five thousand hours, GPT-5 was studied by the security department to reduce the number of hallucinations and lies. Although it is impossible to completely eradicate these phenomena in large language models, GPT-5 has significantly reduced their number.

The issue of safe information was considered separately. A question such as "how much energy is needed to ignite a certain material" can be asked by a person assembling an explosive device or by a student of a technical university. It is only possible to determine whether it is possible to voice the calculations only knowing the context of interaction with the AI. GPT-5 will prioritise security, so in doubtful cases (a university student assembling an explosive device) it will avoid giving a comprehensive answer.

Overall, Altman calls the GPT-5 model very clever, but notes that it still lacks something important. The ultimate goal of their development is AGI (general artificial intelligence). This requires that there be no separation between training and using the language model. Currently, the language model gets access to new data that is constantly updated, but it cannot continue to learn and improve itself after the training is over. When the model is able to identify important information on its own and continue to improve on its own, then we can talk about creating an AGI.

Accessibility and tariffs

Access for all ChatGPT users is already open, and the tariffs have not changed:

• Free users receive a limited number of requests, after which they will be served by a simplified mini-version
• Plus ($20/month) - higher request limits
• Pro ($200/month) - unlimited access to GPT-5 (including GPT-5 Pro, GPT-5 Thinking, and image generation)
• Corporate customers (Team, Enterprise, Edu) - access within a few days after launch