OpenAI has launched GPT-4.5, the successor to its previous GPT-4 technology, marking a significant evolution in its chatbot system. This release is notable for being the last version before a shift to “chain-of-thought reasoning,” requiring more thoughtful interactions rather than immediate responses.
OpenAI rolls out GPT-4.5 but it’s not a frontier model
GPT-4.5, which powers the most expensive version of ChatGPT, may not generate the same level of excitement as its predecessor, GPT-4, due to changes in AI research direction. However, OpenAI asserts that this new technology will provide a more natural conversational experience.
Mia Glaese, vice president of research at OpenAI, emphasized the model’s improved ability to engage in warm, intuitive, and naturally flowing conversations, suggesting a stronger understanding of user intent.
This advancement follows the introduction of OpenAI o1, a system intended to reason through complex tasks such as mathematics, coding, and science. Companies like Google, Meta, and the Chinese start-up DeepSeek are pursuing similar technological advancements to enhance problem-solving capabilities in AI.
According to OpenAI, GPT-4.5 features enhanced writing capabilities, better world knowledge, and a refined personality compared to earlier models. It excels in recognizing patterns and making connections, making it suitable for writing, programming, and solving practical problems.
Despite these upgrades, OpenAI has clarified that GPT-4.5 is not a frontier model. The company stated that it improved GPT-4’s computational efficiency by over 10 times but noted the model does not introduce seven new frontier capabilities compared to previous releases. Its performance is regarded as below that of the o1, o3-mini, and deep research on most preparedness evaluations.
GPT-5 is coming: OpenAI just revealed it
Reports indicated that OpenAI employed the o1 reasoning model, codenamed Strawberry, for training GPT-4.5 using synthetic data. The training methods incorporated new supervision techniques alongside traditional approaches such as supervised fine-tuning and reinforcement learning from human feedback, much like those used for GPT-4o.
OpenAI claims that GPT-4.5 experiences fewer hallucinations than GPT-4o and slightly fewer than its o1 model. Raphael Gontijo Lopes, a researcher at OpenAI, stated that the alignment efforts aimed to make the conversations feel warmer, more intuitive, and emotionally nuanced. Human testers evaluating GPT-4.5 against GPT-4o reported its superior performance across nearly all evaluation categories.
OpenAI CEO Sam Altman described GPT-4.5 as a “giant, expensive model” that “won’t crush benchmarks.” The model launched for ChatGPT Pro users and is set to roll out to Plus and Team users shortly, followed by Enterprise and Edu users. GPT-4.5 is also currently available on Microsoft’s Azure AI Foundry platform, alongside new models from Stability, Cohere, and Microsoft.
The underlying system of GPT-4.5 falls under the category of large language models (L.L.M.s), which learn by analyzing vast amounts of text from the internet, including Wikipedia, books, and chat logs. They identify patterns to generate text independently.
To develop reasoning systems, companies subject L.L.M.s to an additional reinforcement learning process, which can last weeks or months. This method helps the models learn the most effective strategies for solving various problems through extensive trial and error.
While some reasoning systems have outperformed standard L.L.M.s on specific standardized tests, experts caution that such assessments may not accurately reflect real-world performance. The new reasoning systems still face challenges, including the potential for inaccuracies and fabrications, known as hallucination.
OpenAI announced that GPT-4.5 would be accessible to anyone subscribed to the ChatGPT Pro service, priced at $200 per month, starting Thursday.
Featured image credit: OpenAI
