OpenAI has officially announced its new language model, GPT-4.5, which will be available to ChatGPT Pro users. According to the developers, this is the most powerful version of artificial intelligence, featuring improved context understanding, reduced false responses, and higher emotional intelligence.
OpenAI has officially announced its new language model, GPT-4.5, which will be available to ChatGPT Pro users. According to the developers, this is the most powerful version of artificial intelligence, featuring improved context understanding, reduced false responses, and higher emotional intelligence.
The new model will now better understand context when processing queries, which, according to the developers, should show better results on complex mathematical questions. And the overall answers have become more natural.
“This is the first model that feels like I’m talking to a thoughtful human. I’ve had a few moments where the AI’s advice was so good that I just sat back in my chair,” OpenAI CEO Sam Altman wrote in his X.
OpenAI says it has trained GPT-4.5 to use new control methods in combination with traditional methods such as supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF), similar to those used for GPT-4o.
GPT-4.5 was built with scalable learning methods in mind, which allowed the model to become even more reliable and accurate. GPT-4.5 gives 24.3% more accurate answers than GPT-4o. The new model is better at following user queries, making it effective for writing texts, coding, analytics, and other tasks.
Additionally, OpenAI claims that the new model shows reduced error rates and better performance on complex queries, with a 24.7% reduction in incorrect responses.
Altman admits that using GPT-4.5 is extremely expensive, so much so that the company is considering dropping its long-term API support. The cost to developers to access the model is $75 per million input tokens (about 750,000 words) and $150 per million output tokens. By comparison, using GPT-4o costs significantly less — $2.50 per million input tokens and $10 per million output tokens.
