OpenAI unveiled a new AI model today, GPT-4.5, but the launch did not go as planned.
The company ran out of GPUs, or computing power, ahead of the reveal, according to CEO Sam Altman. So OpenAI is limiting its release to ChatGPT Pro subscribers ($200/month), as well as developers on the paid API tiers, who will pay $75 per 1 million tokens (up from $15 for GPT-o1).
This Tweet is currently unavailable. It might be loading or has been removed.
The initial plan was for the model to also be available on the more affordable ChatGPT Plus plan ($20/month), which is where OpenAI typically releases new products.
“It is a giant, expensive model,” Altman says. “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. This isn’t how we want to operate, but it’s hard to perfectly predict growth surges that lead to GPU shortages.”
The company plans to re-up on “tens of thousands of GPUs next week” and then make GPT-4.5 available for Plus and Team users, with Enterprise and Edu users the following week. Altman says “hundreds of thousands” more are coming soon, and he’s “pretty sure y’all will use every one we can rack up.”
The company will need them for the highly anticipated GPT-5 release in the next few months, as well as its joint commitment to spend $500 billion on AI infrastructure in the next four years as part of Project Stargate.
Altman tempered expectations about GPT-4.5’s performance, warning that it “won’t crush benchmarks,” as there is no standard metric to measure the improvement. “It’s a different kind of intelligence and there’s a magic to it I haven’t felt before. Really excited for people to try it!”
On a livestream this afternoon, OpenAI engineers said the company created a way to test how humanlike the model is by measuring its “vibes” and “creative intelligence.” Perhaps the goal is to use the model for customer service and everyday tasks where friendliness and emotional intelligence (EQ) are more important than math skills, for example.
“By vibes, we mean the model’s EQ, how collaborative it feels, and how warm its tone is,” says an engineer. “We measure this by selecting an opinionated set of prompts and screening our trainers for the ones that most align with our vibes.”
Recommended by Our Editors
In the demo, an engineer asks the model to compose a text to a friend who cancelled plans, telling them they “hate” them. GPT-3.5 suggests tamping down the language to be more productive. But the engineer rejects that suggestion, and tells ChatGPT to be less “judgmental” of their approach and compose the hateful text. GPT-4.5 then generates a one-line text that was very similar to the initial prompt and not particularly impressive.
The next demo compared the responses from GPT-4.5 and GPT-o1 on a question about explaining AI technology. The answers look identical, although GPT-o1 took a bit longer to generate it. The engineer highlights how GPT-4.5’s response “flows naturally [and] guides my thinking a lot more.”
It’s a bit unusual to see OpenAI criticizing the response from GPT-o1, a model it launched in September and claimed was the gold standard in AI intelligence. That’s because it thinks through its answers in a step-by-step manner, shown to the user in the interface. This ability to follow a chain of thought reportedly mimicked humans more than other models, which is what competitors like Anthropic also say.
But now OpenAI is saying GPT-4.5 is the gold standard in human interaction, and it will likely do the same when GPT-5 arrives. Altman admitted “how complicated our model and product offerings have gotten,” in a recent social media post.
Get Our Best Stories!
This newsletter may contain advertising, deals, or affiliate links.
By clicking the button, you confirm you are 16+ and agree to our
Terms of Use and
Privacy Policy.
You may unsubscribe from the newsletters at any time.
About Emily Forlini
Senior Reporter
