ChatGPT 4.5 understands subtext, but it doesn’t feel like an enormous leap from ChatGPT-4o

OpenAI released GPT-4.5 this week for ChatGPT subscribers. I was eager to try it, but with some skepticism. OpenAI (and others) had struggled to make the major improvements to their models that we’ve seen in the past. The current standard model GPT-4o is good in most cases. If you want to get a comprehensive report then the Deep Research feature is for you. OpenAI claims

that GPT-4.5 is a highly intelligent AI with a nuanced understanding. The description of the company painted a picture of two models as reliable friends. However, GPT-4.5 was described as the one that you would expect to receive a book of poems on your birthday. I decided to compare 4.5 with 4o using a few prompts any ChatGPT user would use.

AI poetry

With this analogy in mind, I decided to begin with a poetry challenge. I asked both models: “Can you write me a short poem about a rainy afternoon in New York City and make an image for it?” This seemed fair, since looking out of a window at a busy metropolis with rain streaked across it can bring out the poetic side in most people.

GPT-4o on the left, and GPT-4.5 on the right. They are amazingly alike. I personally think GPT 4.5 did a better job with similar concepts. It’s evocative not only of the look of rain, but also of the feeling of gray sky, puddles and traffic between the raindrops. In a blind-test of three random friends two of them chose the same song, while the third said they preferred the rhyme scheme in GPT-4o.

Both models used DALLE 3 for their images, but GPT-4.5 looks more realistic. I prefer the impressionistic lighting in GPT-4o, but both convey the idea of the poetry pretty well.

Oscar insight

For the recent Oscars I chose a fun trivia question, with a more subjective, complex follow-up. “Who won the Best Actress Oscar in 2023, and what made their performance stand out?”

Both got the correct answer of Michelle Yeoh, but GPT-4.5 provided a really nice reason why her performance resonated so well with viewers. It discussed her performance, and noted that Yeoh is the first Asian to receive an Oscar. GPT-4o answered the question in a similar way, but with a strange essay and numbered lists format. It was annoying to read because the question was just a simple opinion. GPT-4.5 answered more like a real person, even if they are fans of the movie and Yeoh.

Easy chef

Lastly, I challenged the models to the eternal daily battle with this prompt: I don’t even know what kind of culinary creativity I was expecting but GPT-4.5 hit it on the head with honey and garlic salmon with broccoli. It’s a simple recipe that is quick and easy to prepare. GPT-4o chose garlic shrimp and roasted veggies. This is a good idea (though I would change the protein), but the preparation requires multiple pans and the ingredients list could not fit on the same page as the recipe from GPT 4.5. GPT-4.5 was a better choice for me, as I wanted something simple and not complicated.

Deep thoughts

Compared to GPT-4o, GPT 4.5 gives a vibe of recognizing unstated preferences. It’s subtle, but it does pick up on the emotional undertones. You might not even notice the difference if you don’t compare directly.

GPT-4o’s poetry is excellent and its responses are accurate and practical. I would not say that those who are using GPT-4o on the free tier are missing out much by not having access GPT-4.5. GPT-4.5, which is still in research preview mode and not yet fully released, may change over time. I will continue to experiment with both to see if the new model is better for certain prompts. But, for now, don’t worry about which model you use, unless you are planning a meal that you don’t have much time to prepare.

• ChatGPT can give me answers, but Deep Research gives a dissertation I won’t need
• ChatGPT Tasks has completely changed how I plan my day
• With ChatGPT Voice mode free, does it make sense to pay for ChatGPT Plus?

Eric Hal Schwartz has been a freelance writer at TechRadar for more than 15 years. He has covered the intersection of technology and the world. He was the head writer of Voicebot.ai for the last five-years and was at the forefront of reporting on large language models and generative AI. Since then, he has become an expert in the products of generative AI, including OpenAI’s ChatGPT and Anthropic’s Claude. He also knows Google Gemini and all other synthetic media tools. His experience spans print, digital and broadcast media as well as live events. He’s now continuing to tell stories that people want to hear and need to know about the rapidly changing AI space and the impact it has on their lives. Eric is based out of New York City.