OpenAI Launches GPT-4.5 with Enhanced Performance, Refined Personality
OpenAI has announced the arrival of GPT-4.5, the latest iteration of their expansive AI language model. Currently available as a research preview exclusively for ChatGPT Pro users, GPT-4.5 is celebrated as the “most knowledgeable yet” by OpenAI. This model marks a significant enhancement over its predecessors, even though OpenAI clarifies it is not categorized as a “frontier model.”
Featuring improved writing capabilities, expanded world knowledge, and a “refined personality,” GPT-4.5 is designed for more natural and intuitive user interactions. OpenAI highlights its enhanced pattern recognition and connection-drawing abilities, making it adept at tasks like writing, programming, and practical problem-solving.
Despite its advancements, OpenAI is candid about the limitations of GPT-4.5. While it achieves a 10x increase in computational efficiency compared to GPT-4, it lacks the groundbreaking innovations anticipated from a frontier model. Initial documentation, later revised, suggested that its performance on preparedness evaluations lags behind certain internal models and deep research benchmarks.
The development of GPT-4.5 involved pioneering “new supervision techniques,” alongside conventional strategies such as supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), methods similar to those used for the GPT-4o. A notable improvement in GPT-4.5 is the reduction of hallucinations compared to GPT-4o, with a slight edge over the o1 model.
OpenAI’s researchers have prioritized making GPT-4.5 a “better collaborator,” striving for warmer, more intuitive, and emotionally nuanced dialogues. Human evaluations have indicated that GPT-4.5 outperforms GPT-4o in a variety of categories, highlighting its collaborative potential.
Sam Altman, OpenAI CEO, acknowledges that GPT-4.5 is a “giant, expensive model” that doesn’t aim to “crush benchmarks” but focuses on more practical improvements that users can directly benefit from.
Following its initial rollout to Pro users, GPT-4.5 will soon be accessible to Plus and Team subscribers in the coming week and subsequently to Enterprise and Edu users. It is also available via Microsoft’s Azure AI Foundry platform, where it joins models from Stability, Cohere, and Microsoft itself.
This release is in line with OpenAI’s strategic roadmap, which earlier projected a late February launch for GPT-4.5 and hints at a potential late May release for GPT-5. GPT-5, as outlined by Altman, is expected to be a “system that integrates a lot of our technology,” including OpenAI’s upcoming o3 reasoning model.
The strategic deployment of GPT-4.5, aimed at practical improvements and enhanced collaboration, signifies a pivotal step towards OpenAI’s ultimate vision of artificial general intelligence (AGI). By progressively integrating its diverse language models and reasoning technologies, OpenAI aspires to develop a more versatile and capable AI system that could redefine interactions between humans and machines.
