OpenAI launches GPT-5

Background

Before GPT-5, ChatGPT offered several models with different strengths. GPT-4o was a general-purpose model for everyday questions, and o3 was a reasoning model for harder problems that benefit from extended thinking. In that arrangement, users had to choose between them. The term reasoning here describes a model that spends more computation working through a problem before it answers, which usually makes it slower and more costly per question than a fast model.

The term hallucination comes up often in discussions of these systems. It describes a model stating false information with confidence.

What happened

On 7 August 2025 OpenAI introduced GPT-5. It called GPT-5 its best AI system yet, with state-of-the-art performance across coding, math, writing, health and visual perception.^[1]

OpenAI's system card describes GPT-5 as a unified system with three parts: a fast model for most questions, a deeper reasoning model for harder problems, and a real-time router that chooses between them based on conversation type, complexity, tool needs and explicit intent.^[2]

GPT-5 rolled out the same day to ChatGPT Free, Plus, Pro and Team users, with Enterprise and Edu users following about a week later. It was the first time free users had access to a reasoning model. After they reached their usage caps, free users fell back to GPT-5 mini.^[3]

Key details

The system card maps the new models to their predecessors. gpt-5-main succeeds GPT-4o, gpt-5-main-mini succeeds GPT-4o-mini, gpt-5-thinking succeeds o3, and gpt-5-thinking-pro succeeds o3 Pro.^[2] In the API, OpenAI offered gpt-5, gpt-5-mini and gpt-5-nano.^[1]

A router, in this sense, is the component that decides where a request goes. The system card's stated criteria are conversation type, complexity, tool needs and explicit intent, so the choice is made by the system rather than by the user. Predecessor names also matter when reading older comparisons, because they show which earlier model each new name replaces.

On factual prompts drawn from real ChatGPT conversations, OpenAI reported that gpt-5-main's hallucination rate was 26% smaller than GPT-4o's, and gpt-5-thinking's was 65% smaller than o3's.^[2] These are OpenAI's own measurements.

OpenAI also treated gpt-5-thinking as High capability in the biological and chemical domain under its Preparedness Framework, the company's method for assessing risks from advanced models. It activated the associated safeguards as a precaution.^[2]

Impact

OpenAI said it expected ChatGPT to reach 700 million weekly active users that week.^[3]

GPT-5 changed the product structure as much as the models themselves. The router takes over a choice that users previously made by picking a model. Whether the router's choices suit each user, and how the figures above hold up in independent testing, are questions the sources do not answer.

Because the router can send simpler requests to a faster model, its effect on speed and cost depends on how often requests go to each. The sources cited here do not report those patterns.