—Our smartest, fastest, most useful model yet, with built-in thinking that puts expert-level intelligence in everyone’s hands, says OpenAI on their launch post.
All users, including the Free tiers should be getting new models directly after launch, while Enterprise and Edu users will have to wait a week.
It will simply be «GPT 5» in the model selector from now on, with the option to use GPT 5-thinking. This can also be triggered by simply writing «think hard about this» or something of the kind directly in the chat.
Introducing GPT-5
Our best AI system yet, rolling out to all ChatGPT users and developers starting today.https://t.co/14Hf1Av6n6
— OpenAI (@OpenAI) August 7, 2025
It’s «the first that it really feels like talking to a PhD-level expert,» OpenAI CEO Sam Altman tells The Verge.
Very nice benchmarks
This is of course carried through in the benchmarks, where GPT-5 Pro is miles ahead in Humanity’s Last Exam, scoring 42% to Grok 4’s 25%.
On AIME 2025, the Pro model scores a perfect 100%, while it clocks in at 32.1% on FrontierMath, and 74.9% on SWE-bench Verified. It’s also the new leader in all categories on LMArena.
First impressions
Initial impressions from using the bot give a feeling of it being snappier, sharper, and less chatty.
It has a cutoff time (the conclusion of its training) of June 2024, but has no problem naming the president that won the election five months later.
Answers also come in faster and more concise, and has stopped answering queries with more questions.
Hardly hallucinates
OpenAI has been mostly concerned with just that real world use, saying «We’ve made significant advances in reducing hallucinations, improving instruction following, and minimizing sycophancy.»
Sycophancy is way down in GPT-5, according to OpenAI, honesty is up – and hallucinations are so rare, it threatens to be a thing of the past.
OpenAI says «When reasoning, GPT‑5 more accurately recognizes when tasks can’t be completed and communicates its limits clearly» — leaving behind a key source of hallucination; when models in the past couldn’t solve a problem, they tended to return a false, but confident answer.
Usage tiers/availability
The difference between the subscription tiers for GPT-5 is the usage volume. Free users get fewer responses, before going down a tier to the new GPT-5 mini model. Plus users can «use it comfortably as their default model for everyday questions,» while Pro users get access to the GPT-5 Pro model, designed for the hardest questions and serious research.
It’s pretty cheap, too, clocking in at $1.25 per million tokens, which makes it significantly cheaper than o3-pro and o3 plain, even though it is much more intelligent.
Missed the live stream? You can catch it here.
Read more: OpenAIs launch post, writeups by The Verge, Axios, Ars Technica.