In the news: OpenAI released GPT 4.5, its largest model yet, focused on “natural” communication

On February 27, OpenAI unveiled GPT 4.5, internally known as “Orion”. The company’s white paper claims that GPT 4.5 is the company’s largest model that has undergone much more resource-intensive training than any of its predecessors. 

Notably, in the original white paper, OpenAI noted it “does not consider GPT 4.5 to be a frontier model.” This point has been later been removed from the latest version of the new white paper. 

GPT 4.5 release in the news

Most notable tech media covered the launch of GPT 4.5. Below is the recap of how the release was commented on in the media lens.

TechCrunch

TechCrunch offers a sober take on the performance of the model. In his piece on the launch, Maxwell Zeff acknowledges the performance increase GPT 4.5 has seen compared to GPT 4o, yet points out a capability the new model does not support yet: a realistic two-voice mode available in GPT 4o. 

TechCrunch notes that OpenAI did not list its hallmark reasoning model, Deep Research, on SimpleQA, which measures how well LLMs answer short, pointed questions. GPT 4.5. was listed on SimpleQA and outperformed GPT 4o. 

CNBC

In a story in GPT 4.5, CNBC reminds readers that GPT 4.5 is particularly significant for OpenAI as the company is in discussions on a $40 billion funding round that would bring the value of OpenAI to $340 billion. 

CNBC quotes Alex Paino, Research Lead at OpenAI, to confirm that the company plans to roll out GPT 4.5 (currently available as a research preview )to Plus and Teams subscribers. 

Reuters

While Reuters covered the GPT 4.5 in a concise write-up, the piece highlights a few details that other media glossed over. 

For example, the article quotes the model’s “hallucination rate” score, 37.1%, compared to the 61.8% GPT 4o has shown and 44% for o1. 

Axios

Axios, as a technology-focused media, gave one of the best reviews of the model written by news outlets. Ina Freid, an Axios reporter, believes that GPT 4.5. can be “the last of its kind”, which means that it will be the last “non-chain-of-thought” model that keeps its reasoning hidden. 

The article explains the drawbacks of the “scaling law” tech companies use to build high-performance models. 

While in the AI world, more seems to be better, Axios points out that “building and powering the massive data centers required to build and run the latest models has become an enormous burden, while assembling ever-bigger datasets has become challenging.”

The New York Times 

The New York Times presents the release of GPT 4.5. In layman’s terms, this helps non-technical readers better understand why the new model is less impactful than the previous one (GPT 4). The article briefly examines reinforcement learning, explaining how it helps train large language models. 

Financial Times

Catering to a financial audience, FT covered the GPT 4.5. release from a cost-benefit perspective. 

Their article on the release explains that while, on the one hand, OpenAI still maintains a superior position on the AI market and is headed towards a $300 billion valuation, on the other hand, the financial burdens of training models like GPT4 and GPT 4.5 and the increasing market competition (most notably from DeepSeek and Anthropic) put Sam Altman’s team under a lot of pressure. 

Key ideas

After diving deeper into the news cycle on GPT 4.5, examining model reviews by industry experts and the model’s original and updated system cards, let’s summarize key noteworthy points about the release. 

• The release was essential to respond to the peer pressure. The timing for GPT 4.5 is all but accidental, and the release has been prompted as a response to Elon Musk, Anthropic, and DeepSeek releasing state-of-the-art LLMs. Some hypotheses say that GPT 4.5. has been around for a long time internally (as its training was cut off back in 2023), and OpenAI did not plan to release it publicly. Given that GPT 4.5. offers no significant benchmarks, these may be highly plausible.

• OpenAI is focusing on non-tech end-users. While its competitors are focusing primarily on the developer audience and target benchmarks significant to engineers (API costs, coding skills, etc.), Sam Altman’s team is focused on offering customers a better experience of interacting with ChatGPT. Hence, it placed its bets on “emotional intelligence” over tech superiority.

• The scaling law is at stake. While performance gains are highly correlated with increasing the degree of LLM scaling, there’s only so far big tech companies can go. The amount of computing power needed to train large language models puts the training costs at hundreds of millions of dollars. Besides, companies have practically run out of data to use for training and need to generate synthetic datasets to mitigate the challenge. The release of GPT 4.5. shows that AI frontrunners will have to make tradeoffs in the direction in which they scale their models. OpenAI focused on improving the personality of its model but lost in technical performance and API costs. 

• The GPT 4.5 release is primarily experimental. With high API costs, OpenAI does not expect GPT 4.5 to gain traction among developers. Still, the company wants to test if its newest model can empower entirely new use cases or do better at creative work most end users typically use it for. 

This was our review of news takes on GPT 4.5 release and its impact on the industry. For more comments on recent AI news and trends, follow Xenoss Blog and our brand page on LinkedIn.