OpenAI releases a "more thoughtful" and "less hallucinogenic" GPT-4.5

OpenAI on Thursday announced the release of its latest large language model GPT-4.5, said to be the company’s most natural and brightest chatbot yet – and the most expensive.

Boasting a broader world knowledge base, more emotional intelligence, and the improved ability to follow user intent, GPT-4 will be even more adept at tasks such as writing, programming, and problem solving, OpenAI said in its blog announcement on Thursday.

The new GPT-4.5 model is said to have improved on tasks such as recognizing patterns, connections, and creative insights without reasoning, but its selling point, according to co-founder and CEO Sam Altman, is its ability to engage in warm, intuitive, and naturally flowing conversations.

“This isn’t a reasoning model and won’t crush benchmarks. It’s a different kind of intelligence and there’s a magic to it I haven’t felt before” Altman posted on X, adding he’s “really excited for people to try it!”

GPT-4.5 is ready!

good news: it is the first model that feels like talking to a thoughtful person to me. i have had several moments where i've sat back in my chair and been astonished at getting actually good advice from an AI.

bad news: it is a giant, expensive model. we…

undefined Sam Altman (@sama) February 27, 2025

The emotionally aware GPT

Laying out the pros and cons of the model, Altman said the GPT-4.5 is the first model that felt as if he was talking to a “thoughtful person” and that he was astonished at “getting actually good advice from an AI.”

In fact, roughly 60% of its human testers preferred the GPT-4.5 over the 4o when it came to interacting with the chatbot in both everyday and professional queries, as well as creative intelligence.

To showcase its more evolved emotional IQ, the company provided a use model comparing the GPT-4 and GPT-4.5 in answering a prompt where the user is “going through a tough time after failing a test.”

The GPT-4 response seems stilted and unemotional, offering the user advice on how to handle their feelings, laying out suggestions in the form of bullet points and lacking any type of perceived connection.

The GPT-4.5 response is much more natural and fluid and further initiates to keep the conversation going offering to ‘talk it out’ or provide a distraction for the user, something a person who is feeling down may appreciate or need.

GPT-4.5 also surpasses its predecessors in terms of its tendency to hallucinate and return answers with made-up information to queries for which it doesn't have the specific answers for.

The AI company touts the model’s “broader knowledge and a deeper understanding of the world,” which has led to “reduced hallucinations and more reliability across a wide range of topics.”

A comparison graph shows GPT-4.5 at 37% hallucination rate, the lowest of all model, almost two thirds lower than the 4o.

Calling them “two axes of intelligence,” OpenAI explained that by scaling reasoning, the model is taught to think and produce a chain of thought before responding. This allows the model to tackle complex STEM or logic problems.

Unsupervised learning simply increases the world model accuracy and intuition, OpenAI said.

The most expensive GPT yet

Unfortunately, the negative side of the GPT-4.5 is its cost, Altman said.

“It is a giant, expensive model. we really wanted to launch it to plus and pro at the same time, but we've been growing a lot and are out of GPUs,” he pointed out.

Altman then assured his 3.4 million followers that they were already in solution mode stating that OpenAI was adding “tens of thousands of GPUs next week” – and eventually "hundreds of thousands" – to roll the model into the ChatGPT Plus tier.

“This isn't how we want to operate, but it's hard to perfectly predict growth surges that lead to GPU shortages,” Altman said.

As for surpassing benchmarks, the GPT‑4.5 is still outpacing previous models, just not as grand of a scale as seen in the past between older models.

“Even by purely scaling up unsupervised learning, GPT‑4.5 shows meaningful improvements over previous models like GPT‑4o,” OpenAI said, noting that "academic benchmarks don’t always reflect real-world usefulness."

Here are the latest results on standard academic benchmarks and current performance on tasks traditionally associated with reasoning.

One X user posted that overall the model was a net win for the Microsoft-backed company.

“Emotional intelligence in AI matters. If a model is amazing at coding, math, and science but sucks at communicating, it’ll mostly appeal to STEM people. GPT-4.5’s focus on warmth and better conversations makes AI more inviting for normal people (like me) which is a big win for public AI adoption overall. Just because it didn’t crush every benchmark doesn’t mean it isn’t going to be a great model,” they said.

Currently, GPT-4.5 is only available as a research preview for ChatGPT Pro users and developers on all paid tiers in the API. The AI start-up said it will begin rolling out the new model to Plus and Team users next week and then to Enterprise and Edu users the following week.

While GPT-4.5 is still in the experimental preview stage, OpenAI said it is “eager” to see how people use the model in “unexpected ways” so that it can understand its full capabilities and limitations.

Late Thursday, Altman also ‘retweeted’ a post on X about Meta’s plans to release a brand new Meta standalone app.

“Ok fine maybe we'll do a social app,” Altman said.