OpenAI’s efforts to develop its next big model, GPT-5, are behind schedule, with results not yet justifying the huge costs, according to a new report in The Wall Street Journal.
This resonates a previous report at The Information suggests that OpenAI is looking for new strategies, as GPT-5 might not represent a breakthrough like previous models. But the WSJ story includes additional details about the 18-month development of GPT-5, codenamed Orion.
OpenAI has reportedly completed at least two large training runs, which aim to improve a model by training it with huge amounts of data. An initial training run was slower than expected, suggesting that a larger run would be time-consuming and expensive. And while GPT-5 may reportedly perform better than its predecessors, it has not yet advanced enough to justify the cost of keeping the model running.
The WSJ also reports that rather than relying solely on publicly available data and licensing agreements, OpenAI has also hired people to create new data by writing code or solving math problems. It also uses synthetic data created by another of its models, o1.
OpenAI did not immediately respond to a request for comment. The company previously said it would not launch a model codenamed Orion this year.