OpenAI released a new base model on Thursday called GPT-4.5, which the company said is its best and smartest model for chat yet. It’s not a reasoning model like OpenAI’s o1 and o3 models, but it can be used to train other models to be reasoning models. Notably, GPT-4.5 was trained using 10 times the computing power (scores of GPUs in data centers) than its predecessor, GPT-4o.
The result is a model whose outputs feel more natural and human, OpenAI said in its press release, and demonstrate a better general understanding of the world. Its writing and programming skills are better, and it hallucinates less. It also displays a higher level of emotional intelligence about the user and what they’re trying to do. For example, when prompted with “I’m going through a tough time after failing a test,” the model responded:
“Aw, I’m really sorry to hear that. Failing a test can feel pretty tough and discouraging, but remember, it’s just one moment—not a reflection of who you are or your capabilities.”
OpenAI cofounder and former researcher Andrej Karpathy, who got early access to the model, posted on X that the improvements are subtle but meaningful. “Everything is a little bit better and it’s awesome, but also not exactly in ways that are trivial to point to,” Karpathy continued; also, the model appears to have improved by 20% in everything it does rather than improving by several times over in certain domains or skills.
OpenAI said GPT-4.5’s performance proves that supersizing models, training data, and computing power can still produce significant performance gains. However, a debate has been raging over this assertion on X.
Karpathy saw evidence of this in his tests: “[I]t is incredibly interesting and exciting as another qualitative measurement of a certain slope of capability that comes ‘for free’ from just pretraining a bigger model.” Of course, “free” is stretching it: Training costs for a model as big as GPT-4.5 could approach $1 billion.
OpenAI is releasing GPT-4.5 as a research preview to ChatGPT Pro users and to developers who pay to access OpenAI models through an API. It will become available to ChatGPT Plus and Team users next week, the company says.
