OpenAI releases GPT-4.5 with improved writing and reduced hallucinations

Today we were introduced to a new language model of ChatGPT, which takes the name of GPT-4.5. The company, in a quick broadcast on its channel on YouTube, highlighted that the new product is its most advanced model to date and that it is up to ten times faster than its predecessor, the GPT 4. Let's go through each of the new features right now.

News in writing

Many people use the ChatGPT to send messages to your loved ones and, if it notices that any word in the text command may express a feeling, such as anger or frustration, the artificial intelligence will use this information to send a more assertive response. OpenAI It also mentions that all responses generated with the support of this new language model have “greater personality”, something that will help users receive responses according to the commands they send.

For this new feature to work, new training methods were used to create the model launched today, such as unsupervised learning. The company explains that the AI ​​learns on its own, without needing ready-made examples. It receives a bunch of messy data and tries to find patterns, organize similar information into groups or simplify what is complex. This is like throwing several pieces of a puzzle on the table and letting the computer figure out how they connect on its own.

This method is used to segment customers, detect fraud or even group similar news. Algorithms such as K-Means (which brings together similar data), PCA (which simplifies information) and neural networks autoencoders help with this. In the end, AI doesn't get the right answers, but it learns to identify what makes sense within the data.

To deliver better responses, the company also opted for scale reasoning. In this method, the goal is to understand how changes in size or quantity affect a system. Something that works on a small scale may not work the same way on a large scale. This applies to everything from physics to artificial intelligence, where a small model can have challenges when applied to large volumes of data.

The result is a model that has broader knowledge and a deeper understanding of the world, leading to reduced hallucinations and more reliability across a wider range of topics.

OpenAI on its new language model GPT-4.5.

Minor hallucinations

Another new feature of the new language model is that it makes fewer mistakes. In theory, “hallucinations"are the phenomena in which artificial intelligences make mistakes and defend the wrong information until the end. Those who simply choose to use the information sent by AIs can end up making serious mistakes, so it is always a good idea to do a fact check before publishing this anywhere.

O GPT-4.5 reaches a rate of 37,5% of hallucinations, while its predecessor reaches 61%. The jump in error rate of the new language model presented today is even greater when compared to the GPT o3 Mini, which has hallucination rates that reach 80,3%. Thus, we have the best language model so far. OpenAI yet.

But it's worth remembering that this is basically an update that paves the way for GPT-5. Still, the GPT-4.5 presents a computational performance ten times greater than that of GPT-4, even though it lags behind experimental models GPT-o1, o3-mini e deep research.

How to use

Anyone who plans to use the new language model can make the switch as soon as they access the home page of ChatGPT. The idea here is that users choose which language model they want to use. However, you can still revert to the previous version if you prefer.

A OpenAI also stressed that all commands for the new language command must be sent in written format, as it does not yet support voice recognition. However, this may be released in the not-so-distant future.

GPT-4.5 has access to the latest and most up-to-date information with search, supports file and image uploads, and can use Canvas to work on writing and coding. However, GPT-4.5 does not currently support multimodal features like Voice Mode, video, and screen sharing in ChatGPT. In the future, we will work to simplify the user experience so that AI “just works” for you.

OpenAI on its new language model GPT-4.5.

Price and availability

The new language model will be released starting next week for plan subscribers. Plus e Our Team. However, subscribers to the plan Pro you can now use the GPT-4.5 and check out what's new from today, whether via desktop, web or smartphone app. The plans of OpenAI have the following prices:

• GPT Plus – R$ 116/month
• GPT Team – R$ 727/month per user
• GPT Pro – R$ 1.164/month

Plan users Enterprise e Edu will receive access to the new language model in the second week of March. While many people were waiting for the GPT-5 for today, in fact, it should only be presented in May, with a system that will integrate several technologies from OpenAI, including the reasoning model o3, announced in December. We will follow all the news closely.

What did you think of GPT-4.5? tell us Comment!

See also other features

With information: OpenAI

Text review done by Jaime Ninece in 28 / 02 / 2025.