OpenAI launches GPT-5 with automatic reasoning

OpenAI has launched GPT-5, the new model that becomes the default ChatGPT model for signed-in users. The main change is not just that it answers better: it now decides automatically when to respond quickly and when to spend more time reasoning.

The rollout began on August 7, 2025 for Free, Plus, Pro, and Team users. Enterprise and Edu users will get access the following week. GPT-5 replaces GPT-4o, o3, o4-mini, GPT-4.1, and GPT-4.5 in ChatGPT.

One system with multiple reasoning levels

GPT-5 combines three components that work together:

• A fast model for most questions.
• GPT-5 thinking, which analyzes difficult problems for longer.
• A routing system that decides which one to use based on the conversation’s complexity, the tools required, and what you ask for.

If you write something like “think carefully about this,” you can force reasoning mode when it is available. The idea is that you should not have to choose manually between several models every time you open ChatGPT.

When usage limits are reached, the system switches to mini versions, which are smaller and faster. For Free users, some reasoning features may take a few days to roll out completely.

More capacity for writing, coding, and understanding images

OpenAI presents GPT-5 as its most capable model yet for everyday and professional tasks. In coding, the company says it is especially better at creating interfaces, applications, and games from a single prompt, as well as finding errors in large code repositories.

In practice, you could describe a web page you need and receive a working first version without having to explain every technical detail. Even so, the result should be reviewed, especially if the code will be used in a real product.

Writing also improves. The model can turn disorganized ideas into reports, emails, or texts with a clearer structure, and follow style instructions more precisely. With images, it better understands charts, presentation photos, diagrams, and other content that is not just text.

The figures highlighted by OpenAI

According to the company’s published evaluations, GPT-5 achieves these results:

• 94.6% on AIME 2025, a mathematics test, without external tools.
• 74.9% on SWE-bench Verified, a coding evaluation using real-world projects.
• 84.2% on MMMU, a multimodal understanding test.
• 46.2% on HealthBench Hard, an evaluation of health-related responses.

OpenAI also says that GPT-5 with reasoning achieves better results than o3 while using between 50% and 80% fewer tokens, meaning less internal text to reach an answer.

These figures come from controlled OpenAI tests and do not guarantee that every daily response will be correct. Their value is in showing the direction of improvement, not in replacing human review.

Fewer errors, less sycophancy

One of the most important changes concerns reliability. In tests using web searches and questions representative of ChatGPT use, OpenAI says GPT-5 was approximately 45% less likely to make factual errors than GPT-4o. In reasoning mode, it was about 80% less likely than o3.

The model also tries to better recognize when it lacks enough information or when a task is impossible. In a test involving deleted images, o3 confidently responded about images that did not exist 86.7% of the time, compared with 9% for GPT-5, according to OpenAI.

GPT-5 also reduces excessively agreeable responses. The company says that answers designed to provoke praise or automatic agreement fell from 14.5% to less than 6%. That could make ChatGPT less flattering and more willing to point out problems in an idea.

What changes for you

GPT-5 should be more useful when a task requires several steps: preparing a report from documents, debugging code, interpreting a chart, or comparing options against specific criteria. The system can coordinate its tools more effectively and follow long instructions without losing as much context.

In healthcare, OpenAI says the model asks more questions, detects potential warning signs, and adapts explanations to the user’s knowledge and location. It does not replace a medical professional: it helps you understand results, prepare for an appointment, and organize your questions.

Access depends on the plan. Free users have usage limits and will switch to GPT-5 mini when they reach them. Plus offers a higher volume, while Pro includes unlimited access to GPT-5 and GPT-5 pro, a variant that spends more time reasoning to produce more complete answers. In tests involving more than 1,000 professional tasks, experts preferred GPT-5 pro over GPT-5 thinking 67.8% of the time.

The next point to watch will not only be how much GPT-5 knows, but when it decides to use its reasoning, which tools it can handle, and how it responds when it does not have a reliable answer. OpenAI’s bet is that AI will stop forcing you to choose between speed and depth, although its limitations still make human oversight necessary for important decisions.