OpenAI released a research preview of GPT-4.5 artificial intelligence (AI) model on Thursday. The successor to the GPT-4o model arrives with improvements in natural conversations, reasoning, and coding. Calling it “our largest and best model for chat yet,” the AI firm highlighted that the company scaled up both pre-training and post-training with this model. The large language model (LLM) was developed using new techniques such as unsupervised learning alongside traditional 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 AI firm announced the release of GPT-4.5. The model is currently available as a research preview to help OpenAI better understand its strengths and weaknesses. Currently, only ChatGPT Pro subscribers have access to the LLM, however, the company said Plus and Team users will get it by next week. Enterprise and Edu users will likely get it after that.
GPT 4.5 benchmark scores
Photo Credit: OpenAI
While GPT-4.5 currently supports search, file and image uploads, and Canvas, it will not support multimodal features such as Voice Mode, real-time video, and screensharing in ChatGPT. For developers, the company is previewing the new AI model in the Chat Completions application programming interface (API), Assistants API, and Batch API on all paid usage tiers.
Interestingly, GPT-4.5 is not a frontier model that surpasses older models in all metrics. The company’s internal benchmark testing shows that the new model is better than the o3-mini in the MMMLU (multilingual), and a couple of coding-related benchmarks. OpenAI CEO Sam Altman said in a post on X (formerly known as Twitter), “This isn’t a reasoning model and won’t crush benchmarks. it’s a different kind of intelligence.”
Comparing responses between GPT 4o and GPT 4.5
Photo Credit: OpenAI
What the CEO meant was that the main focus with GPT 4.5 was improving the AI’s conversational capabilities and nuance understanding. The company claimed that the model has “greater understanding of human needs and intent” and it now responds to conversational prompts in a more human-like manner.
Its creative writing has also improved with a broader knowledge base. GPT-4.5 can follow user intent closely and is said to come with greater emotional quotient (EQ) that also enables it to deliver better output in writing, programming, and solving practical problems. It is also claimed to be hallucinating less than previous models.
Additionally, GPT-4.5 also makes improvements in reasoning capabilities. Notably, it is not a reasoning model and does not rely on test time compute, which increases the processing time to improve the accuracy and depth of a response. Instead, OpenAI said the LLM shows enhanced reasoning as a core capability and can offer better responses in real-time.
These capabilities were said to be added by scaling existing techniques and implementing new techniques in both the pre-training and post-training stages. OpenAI used unsupervised learning, which increased its world model accuracy and intuition. Additionally, SFT and RLHF techniques were also scaled with this model.
OpenAI released a research preview of GPT-4.5 artificial intelligence (AI) model on Thursday. The successor to the GPT-4o model arrives with improvements in natural conversations, reasoning, and coding. Calling it “our largest and best model for chat yet,” the AI firm highlighted that the company scaled up both pre-training and post-training with this model. The large language model (LLM) was developed using new techniques such as unsupervised learning alongside traditional 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 AI firm announced the release of GPT-4.5. The model is currently available as a research preview to help OpenAI better understand its strengths and weaknesses. Currently, only ChatGPT Pro subscribers have access to the LLM, however, the company said Plus and Team users will get it by next week. Enterprise and Edu users will likely get it after that.
GPT 4.5 benchmark scores
Photo Credit: OpenAI
While GPT-4.5 currently supports search, file and image uploads, and Canvas, it will not support multimodal features such as Voice Mode, real-time video, and screensharing in ChatGPT. For developers, the company is previewing the new AI model in the Chat Completions application programming interface (API), Assistants API, and Batch API on all paid usage tiers.
Interestingly, GPT-4.5 is not a frontier model that surpasses older models in all metrics. The company’s internal benchmark testing shows that the new model is better than the o3-mini in the MMMLU (multilingual), and a couple of coding-related benchmarks. OpenAI CEO Sam Altman said in a post on X (formerly known as Twitter), “This isn’t a reasoning model and won’t crush benchmarks. it’s a different kind of intelligence.”
Comparing responses between GPT 4o and GPT 4.5
Photo Credit: OpenAI
What the CEO meant was that the main focus with GPT 4.5 was improving the AI’s conversational capabilities and nuance understanding. The company claimed that the model has “greater understanding of human needs and intent” and it now responds to conversational prompts in a more human-like manner.
Its creative writing has also improved with a broader knowledge base. GPT-4.5 can follow user intent closely and is said to come with greater emotional quotient (EQ) that also enables it to deliver better output in writing, programming, and solving practical problems. It is also claimed to be hallucinating less than previous models.
Additionally, GPT-4.5 also makes improvements in reasoning capabilities. Notably, it is not a reasoning model and does not rely on test time compute, which increases the processing time to improve the accuracy and depth of a response. Instead, OpenAI said the LLM shows enhanced reasoning as a core capability and can offer better responses in real-time.
These capabilities were said to be added by scaling existing techniques and implementing new techniques in both the pre-training and post-training stages. OpenAI used unsupervised learning, which increased its world model accuracy and intuition. Additionally, SFT and RLHF techniques were also scaled with this model.
