What is GPT-5? A guide to OpenAI's new AI model

OpenAI has launched GPT-5, a new AI system that brings advanced capabilities to the company's chatbot. The model, released on 7 August 2025, is a significant update for reasoning, coding, writing, and understanding different types of information. GPT-5 is currently the company's most powerful AI system to date, built to be a unified tool for a wide range of tasks.

How does GPT-5 work?

GPT-5 operates as a single, smart system. Unlike older models, where users had to manually switch between different versions for different tasks, GPT-5 has a built-in router that determines the best approach for a user's request. It can use a quick, efficient model for simple questions and a more detailed reasoning model for complex problems, all in the background. OpenAI CEO Sam Altman described the model as providing access to "an entire team of PhD-level experts."

Key features of the new model

GPT-5 introduces several new features that improve its performance across various domains.

Code and software development

The model introduces "vibe coding" abilities, allowing it to create full applications from simple text descriptions. For example, during a demonstration, GPT-5 built a French learning website with interactive games and flashcards from a single prompt. The model also shows strong performance in other coding tasks, with a 74.9% score on the SWE-bench Verified coding benchmark and an 88% success rate on the Aider Polyglot multi-language coding task. It also helps with advanced code debugging and refactoring.

Reasoning and logic

It shows progress in logical thinking and problem-solving. The model scores 94.6% on the AIME 2025 mathematical reasoning test without tools. It also has an 88.4% performance on GPQA Diamond, a set of PhD-level science questions. The model can engage in multi-step reasoning, revise its own conclusions, and justify its outputs, making it suitable for complex analytical work.

Reduced errors

GPT-5 introduces a "safe completions" approach to reduce factual errors. The model shows a significantly reduced rate of factual errors compared to previous models. This makes it a more reliable tool for professional use in fields like healthcare and legal services.

Health and medical capabilities

The model has new standards for health-related applications with a 46.2% performance on the HealthBench Hard benchmark. This makes it a valuable tool for providing information, supporting healthcare professionals with research, and giving health-related guidance to users.

Multimodal understanding

The model processes text, images, and voice inputs seamlessly. It has enhanced visual reasoning abilities for tasks like chart interpretation, document summarisation from images, and analysing scientific diagrams.

Personalisation and customisation

Users can choose from four new personality modes: Cynic (blunt and sarcastic), Robot (mechanical), Listener (conversational and reflective), and Nerd (enthusiastic with detailed trivia). Paid subscribers can also personalise their ChatGPT interface with custom colour themes and chat thread customisation. The model has persistent memory that remembers user preferences and context across different sessions.

Accessing GPT-5

GPT-5 is available to all ChatGPT users, including those on the free tier.

• Free users get limited daily access to GPT-5 and GPT-5 mini.
• ChatGPT Plus provides extended access with higher usage limits.
• ChatGPT Pro offers unlimited access to GPT-5 and a specialised Pro version with extended reasoning capabilities.

For developers, GPT-5 offers three API variants: GPT-5, GPT-5 Mini, and GPT-5 Nano, each optimised for different use cases and costs. The pricing for the standard API is USD 1.25 per 1M input tokens and USD 10 per 1M output tokens for GPT-5.

Limitations to Consider

Even with its new features, GPT-5 has limits. It can have trouble with extremely long documents or large codebases. The quality of its output still depends on clear instructions from the user. The model is not an example of Artificial General Intelligence (AGI) and still relies on pattern recognition rather than true understanding. Also, while hallucinations have been greatly reduced, they are not completely eliminated. Human oversight is still necessary for high-stakes work.