OpenAI’s recurrent model design is colliding with its safety instrumentation

The Information reports that OpenAI is trying to make its next flagship model safe enough to release after adopting loop transformers, also called recurrent depth. Instead of passing through a fixed stack of unique layers once, this architecture can reuse layers across repeated reasoning steps. That offers an attractive way to trade time for capability and memory efficiency—but it may also weaken chain-of-thought monitoring, one of the imperfect tools laboratories use to spot deception, evasion or other unwanted behaviour.

The architecture is not merely a research possibility. Microsoft briefly described GPT-6 Sol and GPT-6.1 Sol as using looped inference in a blog post, according to the report, then removed that disclosure. OpenAI has not publicly explained its implementation. The important distinction is between capability evidence and safety evidence: a model can perform better because it reasons for longer without leaving supervisors an equally legible account of what occurred inside the loop.

OpenAI has reportedly redirected about 10% of its computing capacity—hundreds of megawatts—toward alignment and safety work following recent agent-security incidents. It also says it is finalising contracts with outside safety assessors. Both are meaningful commitments, but neither establishes that the monitoring problem has been solved. Compute allocation measures effort, while third-party review is only as useful as the access, tests and publication rights the assessors receive.

The release decision therefore tests whether safety evaluation can keep pace with architectural change. If the model’s strongest performance depends on a mechanism that makes existing monitors less reliable, the appropriate response is not simply more of the old testing. OpenAI needs evaluations designed around hidden iterative computation, independent replication and explicit evidence about what failures remain detectable before deployment.