OpenAI has introduced a new large language model called ChatGPT-4.5. The new model offers significant improvements compared to previous generations with a scaled training process and advanced artificial intelligence capabilities. So what are these improvements?

Everything about ChatGPT-4.5! What’s new?

OpenAI says that the ChatGPT-4.5 model provides more accurate information, hallucinates less, and offers a more natural interaction with humans. Moreover, it is available worldwide for Pro users and developers from now on.

GPT-4.5 focuses specifically on scaling the unsupervised learning process. This improves information accuracy by making the model’s knowledge of the world more comprehensive. It also allows the model to better understand user intent and generate more intuitive responses.

As with previous models, GPT-4.5 has a large knowledge base. However, this time, thanks to the innovations made in the training process and optimisation techniques, the hallucination rate has been significantly reduced. It gives more successful results especially in tasks such as writing, programming and problem solving.

OpenAI here adopts two basic paradigms of developing artificial intelligence. These are unsupervised learning and reasoning, respectively. GPT-4.5 also scales unsupervised learning to create a more accurate world model.

In logic execution, OpenAI’s other models, OpenAI o1 and OpenAI o3-mini, focus specifically on improving complex thinking capabilities. Unsupervised learning enables the model to expand its knowledge of the world by using big data and computational power. This model, too, has made significant progress in not only having broader knowledge, but also in presenting that knowledge in a more reliable way.

GPT-4.5 has a higher accuracy rate in simple knowledge questions compared to previous models. According to OpenAI’s tests, the accuracy rate of GPT-4.5 is 62.5 per cent, while the previous generation GPT-4o model remains at 38.2 per cent. A great improvement was also observed in terms of hallucination rates. The hallucination rate of the GPT-4.5 is 37.1%, while the GPT-4o is 61.8%.

It is also seen that GPT-4.5 has made significant gains in evaluations made by human testers. It outperformed GPT-4o with 57.0 per cent in daily queries, 63.2 per cent in professional queries and 56.8 per cent in creative intelligence tests.

This new model also gives more intuitive and contextually strong answers in conversations with people. It better understands the users’ intentions and can provide concise but informative answers instead of producing unnecessarily long answers.

For example, when a user says, ‘I remember an old painting of women burning a ship and refusing to continue the voyage,’ GPT-4.5 immediately responds with the correct answer. You think of Claude Lorrain’s 1643 painting ‘The Trojan Women Setting Fire to Their Fleet’. Inspired by Virgil’s Aeneid, it depicts the Trojan women burning their ship in exhaustion and despair. While previous models may give more ambiguous or misleading answers to this type of question, the GPT-4.5 provides more precise and reliable information.

GPT-4.5 was also developed to more effectively collaborate with people and manage natural dialogues. We can say that the empathy level of the model has increased and it can give more intuitive responses.

In addition, by increasing its capacity to learn from human feedback, GPT-4.5 is better able to recognise when to provide information and when to encourage more conversation in a given context. For example, when a user asks about their thoughts on space exploration, GPT-4.5 provides a more concise and clear answer.

GPT-4.5 also includes important innovations in terms of security. OpenAI announced that it has made the model more secure by using Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) techniques. The model was said to have undergone rigorous security testing under OpenAI’s Readiness Framework and trained with new surveillance techniques to prevent misuse.

ChatGPT-4.5 is available for ChatGPT Pro users on web, mobile and desktop platforms as of today. It will be available for Plus and Team users in the coming weeks. Then it will be extended for Enterprise and Education users. In addition, GPT-4.5’s API integration is also available to developers. It will be accessible via Chat Completions API, Assistants API and Batch API.