Introducing OpenAI o1-preview

OpenAI's o1-preview model scored 84 on one of its hardest jailbreaking tests, compared with 22 for GPT-4o, on a 0-100 scale, as part of a new safety training approach that uses the models' reasoning to apply safety rules. The company says it has formalized agreements with the U.S. and U.K. AI Safety Institutes and granted them early access to a research version of the model. OpenAI also released o1-mini, a model it describes as 80% cheaper than o1-preview and particularly effective at coding.