OpenAI races to fix safety issues in loop-transformer flagship before release
OpenAI is working to get its next, unreleased flagship model into a state leaders consider safe to ship, with alignment work and testing infrastructure still outstanding, according to The Information; employees expect one or two new models by the end of 2026, though it is unclear whether the major upgrade will be among them. The model uses looped transformers, or “Recurrent Depth,” the same technique behind OpenAI's current top model Astra: a token passes through the same transformer layers multiple times before the next word is produced, boosting math and coding performance and lowering cost, but pushing part of the model's reasoning into internal numerical representations that reviewers cannot inspect, undercutting chain-of-thought monitoring. OpenAI reportedly capped Astra's loop count to keep its reasoning readable, and Microsoft accidentally confirmed in a since-edited blog post that GPT-6.1 Sol also uses the technique; the new flagship must additionally clear safety evaluations and US government access provisions before reaching outside partners.
If looping erodes chain-of-thought oversight, safety review could gate OpenAI's next flagship release and shape how other labs adopt the same efficiency technique.