Meet GPT-5: OpenAI launches "world's most advanced" AI model — free for all users

OpenAI has officially unveiled its new flagship GPT-5 artificial intelligence model, which will become the basis for the next generation of ChatGPT.

GPT-5 is OpenAI’s first “unified” AI model that combines reasoning capabilities with fast response. While its predecessor, GPT-4, allowed chatbots to offer intelligent answers to a wide range of questions, GPT-5 will help ChatGPT perform a variety of tasks on behalf of users, such as creating software applications or short research reports, or simpler everyday tasks like managing a calendar.

The main features of the new model are: A “switch” between normal and thinking mode, and a significantly lower number of hallucinations (the “thinking” GPT-5 hallucinated 4.8% of the time, while the o3 and GPT-4o showed 22% and 20.6%) From now on, GPT-5 will be available to all ChatGPT users as a standard model.

OpenAI CEO Sam Altman has already boldly declared that GPT-5 is “the best model in the world” and a significant step towards the development of AI that “can outperform humans” — i.e., AGI (general AI).

Since its launch in 2022, ChatGPT has reached 700 million weekly users and processes 1.7 million requests per minute. However, OpenAI has an even more ambitious goal, which is to have its chatbot used by every eighth person on Earth.

What did GPT-5 show in the tests?

According to OpenAI, GPT-5 is the “most advanced” AI in certain areas, slightly ahead of the leading models from Anthropic, Google DeepMind, and Elon Musk’s xAI in key tests.

• У SWE-bench Verified — a test of real coding tasks taken from GitHub — GPT-5 scored 74.9% on the first attempt. For comparison, Claude Opus 4.1 by Anthropic scored 74.5%, and Gemini 2.5 Pro by Google DeepMind — 59.6%.
• On “The Last Test of Humanity” — a complex test that measures the performance of artificial intelligence models in math, humanities, and sciences — the GPT-5 version with advanced thinking (GPT-5 Pro) scored 42%. This is slightly less than xAI was able to achieve with Grok 4 Heavy, which scored 44.4% in the test.
• У GPQA Diamond — PhD-level scientific questions test — GPT-5 pro scored 89.4% on the first attempt, surpassing Claude Opus 4.1, which scored 80.9%, and Grok 4 Heavy, which scored 88.9%.
• In the test HealthBench Hard Hallucinationswhich measures the accuracy of AI models’ responses to healthcare topics, GPT-5 showed hallucinations in only 1.6% of cases. This is much lower than the previous GPT-4o and o3 models, which scored 12.9% and 15.8%, respectively.
• In the benchmark Tau-benchthat measures the model’s agentic ability to perform simulated online tasks, GPT-5 demonstrates mixed results: in the part of the test that measures the ability of artificial intelligence to navigate airline websites, GPT-5 scored 63.5%, slightly behind o3, which had 64.8%; in another part of the test that measures the ability of artificial intelligence to navigate retail websites, GPT-5 scored 81.1%, behind Claude Opus 4.1, which had 82.4%.

In addition, OpenAI claims that GPT-5 is better than other AI models in harder to measure, subjective areas such as creative design and writing, and is “safer” — read, less prone to intrigue and lying than other models. In addition, GPT-5 is adept at distinguishing between attackers trying to abuse ChatGPT and users making harmless requests.

ChatGPT interface improvements and subscription bonuses

From now on, users can choose one of four characters for the AI: Cynic, Robot, Listener, or Nerd, who will adapt ChatGPT responses to their “personality”.

Although GPT-5 is available to everyone, paid users will have their own advantages:

• The $20 Plus tariff has higher limits;
• Owners of the $200 Pro subscription will have no restrictions at all.

Organizations using the OpenAI Team, Edu, and Enterprise plans will have access to GPT-5 as the default model next week. For developers, GPT-5 will be available in the OpenAI API in three sizes — gpt-5, gpt-5-mini, and gpt-5-nano. The basic model will cost $1.25 per million incoming tokens and $10 per million outgoing tokens.

Source: OpenaAI, TechCrunch