OpenAI will reportedly ship at least one model this year, but its flagship still needs alignment work

Flagship may slip

OpenAI is racing to finish its next flagship model. The model relies on the looped transformer technique used in Astra, which lets the model pass through the same layers multiple times to improve performance in areas like math and coding.

Safety gets harder

That looping comes at a cost: the model's internal workings become more complex, making safety monitoring and alignment work harder. According to the report, the flagship's alignment is still not complete.

Microsoft leak

Microsoft accidentally shared details about how the technique works on a public page, exposing information about the approach.

Year-end plan

Whether or not the flagship is ready in time, employees say one or two new models will likely ship by the end of the year.