OpenAI's flagship model uses recurrent Transformer tech, complicating safety monitoring

OpenAI is racing to finish its next flagship model, which uses the same recurrent Transformer technique as Astra. By passing tokens through the same Transformer layers multiple times, it improves math and coding performance and cuts costs, since a smaller model can do more. But the recurrence also moves some "thinking" out of readable text into internal numerical representations that human reviewers can't inspect, weakening chain-of-thought monitoring. OpenAI deliberately limited Astra's recurrence count to keep its chain of thought readable. Employees say one or two new models could ship before year-end, though it's unclear whether the "major upgrade model" in development is among them. The flagship still needs alignment work and stronger testing infrastructure. Under President Trump's executive order on evaluating advanced AI systems, OpenAI agreed to give the U.S. government access up to 30 days before new models reach external partners and longer before public release. Microsoft accidentally leaked how OpenAI's recurrent reasoning works. In a since-deleted technical article, Microsoft described GPT-6.1 Sol as using "recurrent reasoning," where the model makes multiple "passes" over the same question before answering. GPT-6.1 Sol "uses the same base model weights as GPT-6 Sol, but with two reasoning passes instead of three." A person with direct knowledge confirmed the description matches what OpenAI privately told Microsoft. OpenAI chief scientist Jakub Pachocki has acknowledged that chain-of-thought monitoring is becoming "fragile" and "unfortunately moving in a negative direction."