Support Ukraine’s Army
Youtube
Facebook
OpenAI launches GPT-5 — who gets access and when?
OpenAI has released GPT-5, the next generation of its flagship language model, which is immediately available to all ChatGPT users and developers via an API. According to CEO Sam Altman, the leap in quality is so tangible that you won't want to go back, like you did after the first iPhone with a Retina display.
The Verge writes about it.
Everything we know about GPT-5
OpenAI calls GPT-5 its "smartest and fastest" model with the fewest false positives. Altman compared the evolution: GPT-3 "talked like a high school student," GPT-4 "talked like a college student," and the new version, according to him, is like talking to a PhD expert for the first time. Despite ChatGPT's nearly 700 million weekly users, the company admitted that it has lacked a leading "frontline" model in recent times, and now GPT-5 should return OpenAI to the top of the rankings.
In the chat interface, the model is now presented as a single version without a separate "reasoning" version: behind the scenes, its own router works, which connects a reinforced logical model for more complex queries or on the "think hard" command. ChatGPT head Nick Turley says that the "atmosphere" of GPT-5 is felt even by unprepared users.
The model is immediately available to everyone, but for the free level, there is an unspecified limit of queries, after which the router switches the user to a simplified GPT-5 mini. There are three tariffs for API access: GPT-5, GPT-5 mini, and GPT-5 nano. Four new response "temperaments" have also been added — "Cynic", "Robot", "Listener", and "Geek", as well as the ability to change the color of dialogs.
In OpenAI tests, GPT-5 outperformed its competitors on the SWE-Bench, SWE-Lancer, and Aider Polyglot benchmarks, and Altman predicts the era of "software on demand". During the press briefing, post-training leader Yann Dubois had the model create a French language learning website in a matter of seconds, generating hundreds of lines of code and a full-fledged frontend that worked immediately.
Security remains a focus: According to security chief Alex Boitl, GPT-5 has been tested for more than 5,000 hours, aiming to minimize "hallucinations" and false claims. A new feature called safe completions allows the model to partially answer potentially risky queries, providing only general information that is not suitable for harm. The model is also better at recognizing when it can't give an accurate answer.
OpenAI doesn't disclose GPT-5's training data, but Altman calls it "generally intelligent" and an important step toward AGI, even if "something significant" is still missing — namely, the ability to continuously learn after deployment.
Read also:
