OpenAI unveiled its flagship model GPT-5.

OpenAI has launched its new flagship AI model, GPT-5, which will form the core of the company’s next generation of ChatGPT and API products. This is the company’s first neural network capable of dynamically choosing between fast response and deep task processing modes, depending on the user’s request.

The GPT-5 model significantly outperforms the previous version across a range of metrics. For example, in the SWE-bench Verified test, which assesses programming skills, GPT-5 scored 74.9% on its first attempt — higher than Claude Opus 4.1 and Gemini 2.5 Pro. In PhD-level tasks from GPQA Diamond, it scored 89.4%, becoming the leader among comparable models. The “hallucination” level in GPT-5 is only 4.8%, which is 4–5 times lower than that of GPT-4o and o3.

OpenAI also notes improvements in more subjective areas—creative writing, design, and code generation for educational web applications. Examples include the creation of simulators and visual explanations of complex scientific concepts. GPT-5 can perform tasks on behalf of the user, such as scheduling or generating research materials.

In the free version of ChatGPT, the model is available with limits, but Plus and Pro users receive increased quotas. Three model options are available via the API — GPT-5, GPT-5 mini, and nano — with the ability to integrate into third-party services.

Although GPT-5 has shown outstanding results in terms of accuracy, creativity, and development, it still lags behind its competitors in some tasks. For example, in Tau-bench, the model demonstrated only comparable performance in simulating online actions, and in e-commerce navigation tasks, it lagged behind Claude Opus 4.1.

Sam Altman called the launch of GPT-5 “a significant step toward artificial general intelligence.” The company claims that the new model is not only more productive and safer, but also more effective at identifying malicious queries.