On Thursday, OpenAI unveiled a research preview of its latest artificial intelligence (AI) model, GPT-4.5. This model is an advancement from the previous GPT-4o iteration, featuring enhancements in areas such as natural conversation, reasoning, and coding. OpenAI referred to GPT-4.5 as “our largest and best model for chat yet,” emphasizing improvements in both pre-training and post-training phases. The development of this large language model (LLM) incorporated innovative approaches, including unsupervised learning, alongside conventional supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF).
OpenAI Released GPT-4.5 AI Model
In a blog post, the San Francisco-based organization announced the launch of GPT-4.5. Currently, the model is available as a research preview, aimed at helping OpenAI assess its strengths and weaknesses. At this moment, the LLM is exclusively accessible to ChatGPT Pro subscribers, with plans to roll it out to Plus and Team users by next week. Users in the Enterprise and Education sectors can expect access thereafter.
GPT 4.5 benchmark scores
Photo Credit: OpenAI
GPT-4.5 supports various features including search, file, and image uploads, along with Canvas. However, it does not include multimodal capabilities such as Voice Mode, real-time video, or screen sharing within ChatGPT. For developers, the company is offering a preview of the new AI model in the Chat Completions application programming interface (API), Assistants API, and Batch API for all paid usage tiers.
Notably, GPT-4.5 does not consistently outperform all prior models across every metric. According to OpenAI’s internal benchmark testing, the new model performs better than the o3-mini in multilingual assessments and on certain coding-related benchmarks. OpenAI CEO Sam Altman clarified in a post on X (formerly Twitter) that the model is not designed solely for reasoning and may not excel in all benchmarks, describing it as “a different kind of intelligence.”
Comparing responses between GPT 4o and GPT 4.5
Photo Credit: OpenAI
Altman emphasized that the focus of GPT-4.5 was to enhance the AI’s conversational abilities and its understanding of nuances. OpenAI claims the model demonstrates a “greater understanding of human needs and intent,” facilitating responses that feel more human-like in conversation.
Moreover, improvements in creative writing capabilities are evident, with a broader knowledge base supporting its functions. GPT-4.5 is designed to closely align with user intent and reportedly has a higher emotional quotient (EQ), enhancing its efficiency in tasks related to writing, programming, and problem-solving. It is also said to produce fewer inaccuracies compared to earlier iterations.
Improvements in reasoning capabilities have also been introduced, although the model does not represent a reasoning-centric design and is not dependent on test-time computation. Instead, OpenAI asserts that enhanced reasoning is an intrinsic feature of the model, allowing for more effective real-time responses.
These advancements stem from scaling existing methodologies and integrating novel techniques throughout both the pre-training and post-training processes. OpenAI amplified the use of unsupervised learning to bolster the model’s accuracy and intuitive performance, while also scaling traditional SFT and RLHF techniques with this iteration.
