Openai has just released GPT-4.5 and says it is his largest and best chat model so far

In contrast to argumentation models such as O1 and O3, which work through the answers step by step, “classic” large language models such as GPT-4.5 spit out the first answer that you can think of. But GPT-4.5 is more general. GPT-4.5 rates 62.5% compared to 38.6% for GPT-4O and 15% for O3-Mini.

In addition, the rate with which the models in this test react with invented answers (referred to as hallucination) were 37.1% for GPT-4.5, 59.8% for GPT-4O and 80.3% O3-mini.

In other benchmarks, including MMLU, a standard test for multimodal voice models was the profits of the previous models from Openaai marginal. And GPT-4.5 is worse than O3 for standard benchmarks of standard sciences and mathematics.

The special charm of GPT-4.5 seems to be in conversation. People who are used by Openaai say they have preferred GPT-44 to GPT-4O for everyday questions, professional inquiries and creative tasks, including poems. (Ryder says it is also great in the old school's old school.)

But after years at the top Openai now has a hard amount. “The focus on emotional intelligence and creativity is cool for niche applications such as writing coaches, brainstorming friends,” says Weneem Alshikh, co -founder and CTO of Writer, a startup that develops large language models for corporate customers.

“But GPT-4.5 feels like a shiny new color layer on the same old car,” he says. “If you throw more calculation and data into a model, it can sound more smoothly, but it is not a game channel.”

“The juice is not worth the energy costs, the complexity and the fact that most users do not notice the difference in everyday use,” he says. “I would rather turn them to efficiency or niche problem solutions than overlook the same recipe.”

“GPT-4.5 is open as you cook something bigger behind closed doors. Until then, it feels like a pit stop. “

Sam Altman said that GPT-4.5 will be the last publication in Openai's Classic Line-up and that GPT-5 will be a hybrid that combines a general large voice model with an argumentation model.

In the meantime, Openai is convinced that his oversized approach still has legs. “Personally, I am very optimistic how I find ways through these bottlenecks and continue to be scaled,” says Ryder. “I think the pattern adjustments in the entire human knowledge is extremely profound and exciting.”