OpenAI releases GPT-4.5 as a research preview of its largest chat model
GPT-4.5 scales unsupervised pre-training for broader knowledge and fewer hallucinations, reaching ChatGPT Pro users first before a wider rollout.
OpenAI has released GPT-4.5 as a research preview, described by the company as its largest chat model to date. This version reaches ChatGPT Pro users first, and is only later expanded to a wider audience.
The novelty is not a multi-step reasoning architecture but the scaling of unsupervised pre-training. This approach gives the model a broader knowledge base, better understanding of context and fewer hallucinations, meaning fewer answers that sound plausible but are factually wrong. For business users, this is a factor that directly affects reliability when the model is used to look up information, draft text or summarise documents.
The practical significance is that reducing hallucinations is often worth more than gaining a few points on a leaderboard, because errors in content sent to customers cost far more than the price of a model call. In exchange, the largest model usually comes with higher inference cost and latency, so it is not the optimal choice for every task.
Because the product carries a research preview label, enterprises should treat this as a trial phase: evaluate on their own data and workflows, follow the rollout roadmap and data policies, and hold off on placing the model in critical steps that require committed stability.
Source: openai.com
