OpenAI to ship one or two new models this year but flagship requires alignment work

Employees say OpenAI will likely release one or two new models by year's end, though the major upgraded flagship model may take longer due to necessary alignment work and testing infrastructure setup. These new models utilize the looped transformer technique, or "Recurrent Depth," which boosts math and coding performance and lowers costs, as seen in Astra. However, this looping makes chain-of-thought monitoring harder because the model's "thinking" moves into internal numerical representations. OpenAI has shifted about ten percent of its total compute toward alignment and safety, and must also secure approval from the US government, which requires access to new models up to 30 days before outside partners and longer before public launch.