Update (October 2026): This review is from March 2023, right after OpenAI released GPT-4 on March 14, 2023. GPT-4 is no longer OpenAI’s newest model: in ChatGPT it was replaced by GPT-4o on April 30, 2025, and GPT-5 followed on August 7, 2025.
ChatGPT 4 is out! In this video I tested GPT-4, the new version of OpenAI’s AI language model, live on launch day. Below you’ll find what OpenAI announced, what you needed to use it, and what happened when I tried it myself in ChatGPT Plus.
Official announcement: OpenAI’s GPT-4 page
In this article
What’s new in GPT-4
- OpenAI calls GPT-4 its most advanced system so far, producing safer and more useful responses.
- GPT-4 can solve difficult problems with greater accuracy, thanks to its broader general knowledge and problem-solving abilities.
- GPT-4 is more creative and collaborative than before. It can generate, edit and iterate with you on creative and technical writing tasks, such as composing songs, writing screenplays or learning your writing style.
- It can handle more than 25,000 words of text, about eight times more than the original ChatGPT (around 3,000 words).
- It understands images. In OpenAI’s demo it explains that if the strings in a photo were cut, the balloons would fly away, and it suggests dishes you can make from a photo of ingredients, like crepes, French toast or an omelette.
- OpenAI says it spent six months making GPT-4 safer and more aligned. In OpenAI’s internal tests, it is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5.
- On a simulated bar exam, GPT-4 scored around the top 10% of test takers, while GPT-3.5 scored around the bottom 10%.
How to get GPT-4 (at launch)
The bad news: GPT-4 was not free. In ChatGPT, it was only available with a paid ChatGPT Plus subscription ($20 per month). I upgraded to Plus during the video to test it.
After upgrading, ChatGPT Plus let me choose between three models: Default (GPT-3.5), Legacy (GPT-3.5) and GPT-4. GPT-4 was limited to 100 messages every 4 hours.
My hands-on test
Image test
I gave GPT-4 a link to a photo and asked what it could see. It answered that, based on the URL, the picture showed a blue rose from an online shop. In other words, it guessed from the text in the link. Image input was part of OpenAI’s demo, but it was not available in ChatGPT at launch yet.
Knowledge cutoff
In OpenAI’s launch video, they say GPT-4 finished training in August 2022, so I expected it to know about newer events than GPT-3.5, whose knowledge ends in September 2021. But when I asked GPT-4 about the latest iPhone, it also said its information ends in September 2021. That is correct: GPT-4 finished training in August 2022, but most of its training data stops in September 2021.
Which version am I using?
When I asked GPT-4 which model it was, it said it was based on GPT-3 and knew nothing about a GPT-4 model, even though GPT-4 was selected. This is normal for language models: they only know what was in their training data, and GPT-4 didn’t exist yet when that data was collected.
My first impression
On launch day it was too early to judge. GPT-4 was selected and running, but it told me it was not GPT-4, and image input wasn’t available yet. Test it yourself and let me know in the comments what you think. You can watch the full live test in the video at the top of this post.
