OpenAI's latest model, GPT-5, is officially here. The company announced the model in a "longer than usual" presentation today. The event, which spanned nearly an hour and a half, ran through some of the specific changes that come with GPT-5, as well as some practical use cases, though some of this information was spoiled in a Thursday morning leak.
Introducing GPT-5
First things first: OpenAI is clearly hyped about GPT-5, as any company with a new product would be. But some of the language here is interesting. OpenAI CEO Sam Altman likened the different AI models, or GPTs, to people of different ages and skill sets: Altman said GPT-3 is like talking to a high school student; GPT-4 is like talking to a college student; and GPT-5 is like talking to an "expert," a keyword thrown around quite a bit during this presentation.
As with any new model, OpenAI says GPT-5 out performs previous GPTs and models on the market across a number of benchmarks, including SWE-bench, Aider Polyglot, MMMU, and AIME 2025. What piqued my interest in particular are OpenAI's claims that GPT-5 has a far lower hallucination rate than previous models. AI models have a bad habit of hallucinating, or, in plain terms, making things up. OpenAI says GPT-5 makes up information less than other models, which will be interesting to put to the test.
Speaking of which, the model will be available to test starting today, Aug. 7. What's more, OpenAI is making the full GPT-5 model available to all users—including those who use ChatGPT for free. Free users do have the lowest use limit when it comes to GPT-5, and when they hit it, they'll be bumped to GPT-5-mini, a lightweight version of the model. Plus users, those who pay $20 per month for ChatGPT, have higher limits than free users, but will still be kicked to the mini model when they hit their limits, while Pro users ($200 per month) have no limits, as well as access to GPT-5 Pro. OpenAI didn't share what those limits are in the presentation, though they may reveal that on their website in a separate post.
One key feature of GPT-5 is its ability to switch between a thinking and non-thinking model depending on the prompt. If you ask ChatGPT something simple, GPT-5 will respond with a quick, but detailed, answer. Ask it something more complex—say, a request to create something—and the thinking model kicks in. You'll then be able to see the model break down the request step by step, and watch as it reasons through its response. If you pay for ChatGPT, you can access the thinking model from the model picker, but all users can request that GPT-5 think during the prompt.
OpenAI demonstrators casually announced that GPT-5 deprecates older GPT models, which means you likely won't be able to use older models like GPT-4o or GPT-4.1 going forward. As such, demonstrators asked both GPT-4o and GPT-5 to write a eulogy for older models, showing how GPT-5's writing is less cookie cutter than 4o and earlier. I disagree, though. While GPT-5 might have a bit more personality and flourish than GPT-4o, its writing had plenty of the telltale signs of AI-generated text, including flowery language and awkward similes.
Models
As with previous GPTs, GPT-5 comes in three different models. These include:
GPT-5: This model is for logic and multi-step tasks.
GPT-5-mini: This model is not as powerful as GPT-5, and, as such, is designed for "cost-sensitive" applications.
GPT-5-nano: This model is designed to be fast for "low latency" applications.
This isn't a huge surprise, since the news was leaked in an accidental GitHub post early Thursday morning.
Coding
My take is that OpenAI is particularly happy with GPT-5's coding abilities. Much of the presentation was spent on demoing how the model can help write code, craft programs from scratch, and hunt for bugs. In the first example, a demonstrator asked the bot to create a website to help their partner learn French. They wanted GPT-5 to create three elements: a flashcard section; a quiz; and a game; specifically a recreation of Snake, but with a mouse hunting cheese, and anytime the mouse reached the cheese, the game would speak a word in French out loud.
Sure enough, GPT-5 generated the site. In fact, the demonstrator had it generate three different versions of the site, as a way of showing the variety of GPT-5 outputs. The sites were fine, if not a big plain—though don't get me wrong, it's incredible an LLM can do this at all, in just minutes. I'm not sure we're anywhere close to being able to describe a high-quality website and expect a chatbot to spit it out in a minute, but this doesn't make me want to invest much time in learning how to code.
In another example, the demonstrator asked ChatGPT to build a balloon popping game set around a castle on a mountain. Instead of waiting for the bot to generate the game, they pulled up a previously generated game, which means you should take this with a grain of salt. But sure enough, OpenAI demoed a crude 3D game which let you shoot balloons around a castle on a hill. You could even "chat" with the soldiers and characters around the castle, though in practice it was more of a simple chatbot function with different personalities.
Deception and safety
OpenAI touched on GPT-5's deceptions, or, in other words, its ability to lie. This is different from hallucinations, as deceptions seem more deliberate—rather than simply making up information, the bot actively attempts to deceive you, perhaps in response to news that you're shutting it down. Scary stuff.
OpenAI claims that GPT-5 is less deceptive than both o3 and o4-mini. Plus, if you ask it a question that would normally be refused by a previous model due to safety concerns, it'll try to answer the question in a impartial or objective way—in an attempt to service those who are asking legitimate questions. If it must refuse, it'll give you a detailed answer as to why.
Health
Sam Altman was particularly keen on hyping up GPT-5's health capabilities. Altman called GPT-5 the "best model ever for health," and that it scores higher than any previous model on OpenAI's HealthBench benchmark.
Altman then brought out a couple to discuss their use of ChatGPT in navigating a difficult and complex cancer diagnosis. While most of their experience was using previous ChatGPT models, the couple did note that GPT-5 was particularly helpful in going deeper on their questions—giving them additional information, including questions they should ask the doctor or what to expect next.
Other new features
Demonstrators also touched on Voice Mode, but not for long. The major demo was of Study and Learn Mode, which supposedly helps you learn subjects using Voice Mode. They demoed this by asking the bot to help them learn Korean, specifically to slow down its speech so they could hear each individual word clearly. The bot did, though it didn't slow down quite enough for me, someone who does not speak Korean, to understand. Once GPT-5 hits my ChatGPT account, I'll be curious to see if you can ask the bot to speak even slower.
Additional, smaller features announced include the ability to customize the color of your ChatGPT conversations, as well as Gmail and Google Calendar integration. The latter gives me privacy concerns, but I imagine any Google Workspace users who also rely on ChatGPT will enjoy this collaboration.
Disclosure: Lifehacker’s parent company, Ziff Davis, filed a lawsuit against OpenAI in April, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.
