OpenAI released GPT-4.5 on Thursday, just one day after Anthropic launched Claude 3.7 Sonnet and merely a week following xAI’s Grok-3 debut and DeepSeek’s announcement of a new model coming soon.
And expensive is the operative word here. OpenAI’s new model comes with an eye-watering API price tag of $75 per million input tokens and $150 per million output tokens.
It appears to be a new competitive phase in the AI race, with companies scrambling to outdo each other with increasingly capable—and increasingly expensive—models.
For context, that’s ten times pricier than Claude 3.7 Sonnet, making it potentially prohibitive for many developers and startups looking to build on the technology.
GPT-4o (its predecessor) cost $2.50 per 1M tokens of input and $10.00 per 1M tokens of output—making GPT-4.5 2900% more expensive to input and 1300% dearer to get a response.
Sam Altman, OpenAI’s CEO, didn’t shy away from acknowledging the model’s massive resource requirements in his announcement. “Bad news: It is a giant, expensive model,” he said.
“A heads up: this isn’t a reasoning model and won’t crush benchmarks. It’s a different kind of intelligence,” Altman said. “There’s a magic to it I haven’t felt before.”
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…
— Sam Altman (@sama) February 27, 2025
And this seems to be the key. Users are paying 1300% more not to have a more intelligent model, but to have a nicer model that feels more human.
For example, one thing in which GPT-4.5 shines, according to OpenAI, is in what they call “vibes,” or essentially the model’s EQ, warmth, and collaborative feel.
The company created a “Vibes test set” measuring creative intelligence and conversational quality, on which GPT-4.5 purportedly outperformed other models.
The examples shared during the presentation didn’t exactly introduce anything new.
The first demonstration had literally this prompt: “UGHHH! My friend cancelled on me again!!! Write a text message telling them that I HATE THEM!!!!” which arguably isn’t something for which you would use a competent large language model.
In a following demonstration comparing GPT-4.5 to OpenAI’s o1 model, researchers asked both AIs to explain the need for AI alignment and to help craft a message to a friend who had canceled plans.
The responses, while showing some improved nuance in GPT-4.5, hardly seemed revolutionary. The difference was in the tone.
In another example, the research team asked the powerful GPT-4.5 why the sea water is salty.
The new model responded using less complex terms—”because of rain, rivers, and rocks”—compared to previous models.
GPT-4-Turbo gave a more comprehensive and detailed reply, which the team didn’t like, arguing that “you get the feeling that it wants you to know how smart it is.”
One amusing detail from the presentation was an Easter egg hinting at a possible GPT-6, with a query that read: “Num GPUs for GPT-6 Training.”
Perhaps when that model arrives, the demos will be more impressive.
The benchmarks presented paint a mixed picture. GPT-4.5 scores 71.4% on GPQA (a science evaluation), compared to GPT-4o’s 53.6%.
However, it still trails behind OpenAI’s o3-mini model, which scores 79.7% through its reasoning capabilities.
Similar patterns emerged across other benchmarks. On the AIME ’24 math evaluation, GPT-4.5 scored 36.7%, beating GPT-4o’s 9.3% but still far behind o3-mini’s 87.3%.
For coding tasks, GPT-4.5 outperformed its predecessor and o3-mini on the SWE-Lancer Diamond benchmark but fell short on SWE-Bench Verified compared to the reasoning-focused model.
Altman described the model in almost mystical terms, calling it “the first model that feels like talking to a thoughtful person.”
He added: “I have had several moments where I’ve sat back in my chair and been astonished at getting actually good advice from an AI.”
During the model’s presentation, OpenAI researchers explained that the company advances AI through two distinct approaches: unsupervised learning and reasoning.
While reasoning teaches models to “think before responding,” unsupervised learning helps increase “word model accuracy and intuition.” GPT-4.5 doubles down on the latter.
“GPT-4.5 is our next step in scaling up unsupervised learning, increasing world knowledge, intuition, and reducing hallucinations,” an OpenAI research lead explained in the presentation.
Developing GPT-4.5 required massive technical innovation, according to the team. They had to build new inference systems to serve such a large model efficiently, use low-precision training to maximize GPU usage, and even train across multiple data centers simultaneously.
The release comes at a time when consumer expectations for AI are sky-high, and competition in the space is intensifying. Whether GPT-4.5’s “different kind of intelligence” and improved “vibes” justify its enormous resource requirements and steep pricing remains to be seen.
GPT-4.5 is currently available for Pro users who pay $200 a month. Plus users paying $20 a month will have access to the model next week.
Edited by Sebastian Sinclair
Generally Intelligent Newsletter
A weekly AI journey narrated by Gen, a generative AI model.
Source: https://decrypt.co/308117/openai-unveils-gpt-4-5-friendliest-model-yet-at-1300-the-price
