OpenAI Unveils GPT-4.5

On Thursday, OpenAI launched its latest model, GPT-4.5, which is positioned as a more versatile alternative to its previous STEM-oriented reasoning models. According to OpenAI, GPT-4.5 excels in various tasks such as writing, programming, and addressing practical problems.

A New Benchmark in AI

OpenAI CEO Sam Altman described GPT-4.5 on his X account as a breakthrough, stating it’s “the first model that feels like having a conversation with a thoughtful person.” He remarked on the model’s ability to provide genuinely useful advice, noting moments where the AI’s responses left him impressed.

Altman further indicated that the model is set to be “massive” and “costly.” Although there was an intention to roll out the model simultaneously for both Plus and Pro subscribers, a shortage of GPUs led to a delay. He assured that “tens of thousands of GPUs” would be made available the following week, allowing for an extended rollout to the Plus tier.

The Scaling Debate Continues

In Silicon Valley, there is ongoing discussion regarding the optimal methods to enhance AI models’ capabilities. GPT-4.5 supports the prevailing belief that increased data and computational power contribute to superior model performance. Mark Chen, OpenAI’s chief research officer, stated in an interview with Big Technology that the company has yet to experience diminishing returns from their scaling efforts. “We make projections based on previous models to predict performance. This is the point that signifies the next order of magnitude,” Chen added.

Despite high training expenses, OpenAI has identified cost-effective methods for operating larger models. Chen pointed out that inference costs have significantly decreased since the original release of GPT-4.

Rolling Out Access

As of Thursday, GPT-4.5 is available in a research preview to users paying $200 per month for ChatGPT Pro, as well as developers accessing it through the API. OpenAI plans to extend access to ChatGPT Plus, Team, and Edu users in the coming week.

During a live demonstration of GPT-4.5, Amelia Glaese from OpenAI’s technical team explained that this model represents the latest iteration of their “unsupervised learning” approach, which prioritizes scaling models based on “word knowledge, intuition, and minimizing hallucinations.”

In terms of interaction, GPT-4.5 showcases a superior capability for engaging in natural conversation and exhibiting emotional intelligence. Compared to its predecessor, the STEM-focused reasoning model o1, GPT-4.5 is more responsive to social cues. Raphael Lopes, also part of OpenAI’s technical team, illustrated how GPT-4.5 could tactfully rephrase an angry message to a friend—demonstrating a higher level of contextual awareness than o1.