GPT-4.5: OpenAI’s Latest and Most Advanced Conversational Model
Introduction to GPT-4.5
GPT-4.5, OpenAI’s main large language model, was just released. It is the company’s largest and best all-around conversation model to date. According to OpenAI research scientist Mia Glaese, “it’s really a step forward for us.”
OpenAI’s Two Product Lines
OpenAI has been promoting two product lines since the release of its so-called reasoning models, o1 and o3. The non-reasoning lineup includes GPT-4.5, which research scientist Nick Ryder, Glaese’s colleague, refers to as “an installment in the classic GPT series.”
GPT-4.5 Availability and Pricing
GPT-4.5 is now available to those with a $200/month ChatGPT Pro membership. According to OpenAI, the rollout to other users will start the following week.
Bigger Models, Better Performance?
Bigger is better, as OpenAI has demonstrated with each release of its GPT models. However, there has been a lot of discussion about how that strategy is failing, including comments made by Ilya Sutskever, the former top scientist at OpenAI. To the skeptics, the company’s assertions on GPT-4.5 seem like a thumb in the eye.
Pattern Recognition and Emotional Intelligence
When trained on billions of pages, all large language models identify patterns. Basic concepts and syntax were taught to smaller models. According to Ryder, “all of these subtle patterns that come through a human conversation—those are the bits that these larger and larger models will pick up on.” This means that larger models are able to identify more specific patterns, such as emotional indicators, such as when a speaker’s words convey animosity.
According to Glaese, “it has the ability to engage in warm, intuitive, natural, flowing conversations.” “And we believe that it has a better comprehension of what users mean, particularly when their expectations are more implicit, resulting in responses that are nuanced and considerate.”
Training and Model Size
At this stage, Ryder adds, “we kind of know what the engine looks like, and now it’s really about making it hum.” “The main goals of this exercise are to increase the compute and data sizes, identify more effective training techniques, and then push the boundaries.”
The precise size of OpenAI’s new model is unknown. However, it states that the scale jump from GPT-3.5 to GPT-4o is equivalent to the scale jump from GPT-4o to GPT-4.5. According to experts, GPT-4 may have up to 1.8 trillion parameters—values that are adjusted during model training.
Training Techniques and Improvements
Similar methods to those employed for GPT-4o, including human-led fine-tuning and reinforcement learning with human input, were used to train GPT-4.5.
Finding scalable paradigms that allow us to invest more and more resources to produce increasingly intelligent systems is the secret to developing intelligent systems, according to Ryder. This is a formula that has been used for many years.
Performance on Benchmarks
Normal big language models like GPT-4.5 spew out the first response they think of, in contrast to reasoning models like o1 and o3, which go through replies step by step. GPT-4.5, however, is more versatile. When tested on SimpleQA, an OpenAI-created general-knowledge test last year that covers everything from science and technology to TV series and video games, GPT-4.5 scores 62.5%, while GPT-4o scores 38.6% and o3-mini scores 15%.
Furthermore, according to OpenAI, GPT-4.5 gives considerably less fictitious responses, or hallucinations. GPT-4.5 made up answers 37.1% of the time on the same test, compared to 80.3% for o3-mini and 59.8% for GPT-4o.
Comparison with Other Models
However, SimpleQA is only one standard. The improvements over OpenAI’s earlier models were only slight on other tests, such as MMLU, a more widely used benchmark for evaluating big language models. Also, GPT-4.5 performs lower than o3 on common math and science tests.
The unique attraction of GPT-4.5 appears to be its dialogue. OpenAI’s human testers report that for both professional and ordinary queries as well as creative tasks like writing poetry, they favored GPT-4.5 over GPT-4o. (Ryder claims that it excels at traditional internet ACSII art as well.)
Competition and Industry Reactions
However, OpenAI is up against a difficult bunch after years at the top. According to Waseem Alshikh, cofounder and CTO of Writer, a firm that creates massive language models for enterprise clients, “the emphasis on emotional intelligence and creativity is cool for niche use cases like writing coaches and brainstorming buddies.”
However, he claims that GPT-4.5 feels like a brand-new coat of paint applied to an old vehicle. “A model may sound smoother if more computation and data are applied, but it won’t change the game.”
Challenges and Criticism
“When you take into account the energy costs and the fact that most users won’t notice the difference in daily use, the juice isn’t worth the squeeze,” he argues. “Rather than continue supersizing the same recipe, I’d prefer to see them shift to efficiency or specialized problem-solving.”
The Future: GPT-5 and Beyond
According to Sam Altman, GPT-4.5 will be the final iteration of OpenAI’s standard lineup, while GPT-5 will be a hybrid that blends a reasoning model with a general-purpose large language model.
“OpenAI is phoning it in with GPT-4.5 while they are working on something more ambitious behind closed doors,” Alshikh argues. “This feels like a stopgap until then.”
Nevertheless, OpenAI maintains that its massive strategy is still viable. “I’m really hopeful that we can figure out how to get past those obstacles and keep growing,” Ryder adds. “I believe that pattern-matching throughout human knowledge has a profound and fascinating quality.”
