GPT-5 is OpenAI's new flagship AI system, released yesterday to every ChatGPT user, including the free tier. It combines a fast model, a deeper "thinking" model and a router that decides which one should answer each message, so most people no longer have to pick a model themselves.

What is GPT-5?

On 7 August 2025 OpenAI introduced GPT-5 and began rolling it out to ChatGPT's Free, Plus, Pro and Team users, with Enterprise and Education accounts to follow next week. It is not just a bigger version of GPT-4o. OpenAI describes GPT-5 as a unified system: a quick, efficient model for most questions, a deeper reasoning model called GPT-5 thinking for harder ones, and a real-time router that chooses between them.

That design answers a long-standing complaint. Until now, ChatGPT offered a menu of models with names like GPT-4o, o3, o4-mini and GPT-4.5, and users had to guess which to use. As The Verge reports, Sam Altman called the old picker a "very confusing mess." GPT-5 replaces all of those as the default for signed-in users.

How the router works

When you send a message, the router looks at the kind of conversation, how complex the request is, whether tools such as search are needed, and any explicit signal from you. If you write "think hard about this," it is more likely to send the request to the thinking model. OpenAI says the router keeps learning from real use, such as when people switch models or which answers they prefer.

Here is how that plays out in practice:

You ask	Likely path
"What's a good synonym for 'happy'?"	Fast model, near-instant reply
"Plan a three-week budget trip through Portugal for a family of four"	May switch to thinking for a more careful answer
"Think hard about this: which of these two job offers is better over five years?"	Thinking model

OpenAI says it plans to merge these abilities into a single model in the future.

What OpenAI says is better

OpenAI's launch post focuses on reliability as much as raw intelligence. It reports that, with web search on, GPT-5's answers contain a factual error roughly 45% less often than GPT-4o's, and that the thinking version makes far fewer errors than o3. The company also says GPT-5 is more honest about tasks it cannot complete. In one test where images were removed from questions about charts, o3 still answered confidently about the missing pictures 86.7% of the time, compared with 9% for GPT-5.

Other changes:

Less flattery. OpenAI says GPT-5 is less likely to simply agree with you, cutting sycophantic replies in its tests from 14.5% to under 6%.
Safer partial answers. Instead of flatly refusing borderline requests, GPT-5 uses what OpenAI calls safe completions: it gives the most helpful answer it safely can, sometimes only at a high level, and explains when and why it won't go further.
Coding and writing. OpenAI calls it its strongest coding model so far, especially for building web front ends and debugging large projects, and its most capable writing partner.
Health questions. OpenAI says it scores higher than previous models on HealthBench, its medical-scenario evaluation, while stressing that ChatGPT does not replace a doctor.

These are the company's own measurements, so it's worth testing them on your own work. Independent evaluations will take a while to appear.

Access and limits

The difference between free and paid plans is mostly how much you can use. Free users get GPT-5 until they hit a usage cap, after which ChatGPT switches to GPT-5 mini, a smaller model; OpenAI notes full reasoning may take a few days to reach every free account. Plus users get much higher limits. Pro subscribers get unlimited GPT-5 plus GPT-5 pro, a version that thinks even longer for the hardest problems. Developers get three sizes in the API: the full GPT-5 plus smaller mini and nano versions.

Try it: Ask ChatGPT the same tricky question twice: once plainly, and once starting with "Think hard about this." Compare how long each took and how the answers differ in depth and caution.

What changes for everyday ChatGPT users?

The biggest change is that you can stop choosing. Open ChatGPT, type your question and GPT-5 decides whether a quick answer will do or whether to reason first. Paid users can still pick "GPT-5 Thinking" from the menu when they want to force careful reasoning, and anyone can nudge the router by asking it to think hard. A good habit is to save that nudge for decisions, calculations and plans, where a slower, more careful answer is worth the wait.

You'll also notice a different tone. OpenAI says GPT-5 uses fewer unnecessary emojis and is less effusive than GPT-4o. If you want a specific style, a research preview adds four preset personalities: Cynic, Robot, Listener and Nerd. They are optional and can be changed in settings at any time, and OpenAI says each still meets its bar for avoiding sycophancy. The Verge adds that you can also pick a color for individual chats.

Your prompting habits mostly still work. Clear context, a specific goal and a description of what a good answer looks like help as much as ever. Because GPT-5 follows instructions more closely, OpenAI says it also respects custom instructions better, so it may be worth revisiting what you've saved there.

Finally, treat the improvements as better odds, not guarantees. Fewer factual errors still means some. For anything that matters, such as health, money or legal questions, ask GPT-5 to show its sources and check them yourself.