Adapted by:
Javier Pastor
Senior WriterComputer scientist turned tech journalist. I've written about almost everything related to technology, but I specialize in hardware, operating systems and cryptocurrencies. I like writing about tech so much that I do it both for Xataka and Incognitosis, my personal blog.
Alba Mora
WriterAn established tech journalist, I entered the world of consumer tech by chance in 2018. In my writing and translating career, I've also covered a diverse range of topics, including entertainment, travel, science, and the economy.
OpenAI CEO Sam Altman had already warned that the company was planning to release GPT-4.5 very soon. Users had been waiting for the successor to GPT-4 for months, but expectations had begun to lower over time. Some experts pointed to the slowing of AI progress and the challenges of scaling. GPT-4.5 was meant to prove them wrong. However, OpenAI’s latest model presents several issues.
Launching GPT-4.5. On Thursday, OpenAI finally unveiled GPT-4.5, the theoretical successor to GPT-4. “It is the first model that feels like talking to a thoughtful person to me,” Altman shared in an X post.
Gigantic and costly. Altman also acknowledged a significant drawback, saying, “Bad news: It is a giant, expensive model.” He mentioned that the company has run out of GPUs for a mass release, so the availability of GPT-4.5 is currently very limited. In fact, only ChatGPT Pro users can access it at the moment.
Click on the image to view the post on X.
Very costly indeed. Using the GPT-4.5 model via the OpenAI API is extraordinarily expensive. It costs $75 per million input tokens and $150 per million output tokens. In comparison, GPT-4 costs $2.50 and $10 respectively–30 to 15 times less. Meanwhile, o1, the most expensive model until now, costs $15 and $60 for input and output tokens respectively.
Not a frontier model. OpenAI’s system card says that, unlike GPT-4, GPT-4.5 isn’t a frontier model. This distinction is important because, although it’s the company’s largest language model, frontier models are generally more capable and large-scale. This can lead to risks such as generating misinformation or bypassing established rules. In GPT-4.5, OpenAI appears to have focused significantly on reducing errors. It seems to be one of its advantages, given that it reportedly makes fewer mistakes according to some test results.
Click on the image to view the post on X.
No better in nearly any aspect. Tests and benchmarks indicate that the performance jump is particularly disappointing, especially compared to its competitors’ newest models. It performs worse than Perplexity’s Deep Research in terms of factual accuracy. According to TechCrunch and several experts, it also lags behind Claude 3.7 Sonnet in coding capabilities. Additionally, it’s inferior in reasoning to models such as DeepSeek R1, o3-mini, and Claude 3.7 Sonnet (which is a “hybrid” model) despite not being primarily focused on that aspect.
Mixed feelings. According to other experts, GPT-4.5 is slow, only upgraded until October 2023, and doesn’t represent a significant advancement. Developer Simon Willison shared an AI-generated summary of several online discussions regarding GPT-4.5. The conclusions are clear. Users believe the model’s naming is inappropriate, and the performance doesn’t meet expectations after a long wait. Many also say it’s too expensive, and the price/performance ratio is highly questionable.
Meanwhile, former Tesla AI director Andrej Karpathy concluded, “Everything is a little bit better and it’s awesome, but also not exactly in ways that are trivial to point to.”
More human? Regarding the model’s conversational abilities, Altman noted that he was surprised by GPT-4.5’s proficiency, which may indicate where this model truly excels. Karpathy also pointed out that improvements in this area might be due to better “creativity, analogy making, general understanding, [and] humor.” All this could make conversations with GPT-4.5 feel even closer to human interactions.
Scaling doesn’t work, the slowdown is apparent. GPT-4.5 exemplifies how the AI sector may have reached the limits of scaling. Developing a gigantic LLM no longer appears to provide clear advantages over its predecessors. Investing more data and GPUs into training these models also seems impractical. Altman has explicitly said that GPT-4.5 will be the company’s last non-reasoning model, signaling that the generative AI slowdown, at least concerning traditional models, is a reality.
Why did OpenAI release GPT-4.5 then? The company says in a press release, “We’re sharing GPT-4.5 as a research preview to better understand its strengths and limitations. We’re still exploring what it’s capable of and are eager to see how people use it in ways we might not have expected.” This reflects the doubts its creators have about the model and raises the question of why they chose to release it.
Creating the hype. OpenAI needs to keep generating buzz around its models, especially given that its competitors are gaining traction. Models such as Claude 3.7, Grok 3, and DeepSeek R1 have turned the tables. They now pose a significant challenge to OpenAI, which recently seemed ahead of the competition. The landscape is now less clear, and competitors are already outperforming OpenAI’s models in many ways. OpenAI needs to assert its presence. However, the release of GPT-4.5 may backfire, given that its performance, at least initially, is somewhat disappointing.
Investor pressure. Some speculate that OpenAI was compelled to release GPT-4.5 to satisfy investors. After all, they’ve poured billions into the company and are probably seeking reassurance about their investment. Unfortunately for OpenAI, it seems GPT-4.5 may not provide that reassurance. With this launch, it’ll likely be challenging for the AI company to attract new investors.
Image | OpenAI
Log in to leave a comment