OpenAI scrambles to fix safety issues on next flagship model

OpenAI is working through safety concerns tied to its next unreleased flagship, The Information reported, as leaders try to get the model to a point where they believe it can be shipped. At the center sits an architectural choice called loop transformers, also known as recurrent depth, where the same block of layers is applied several times in a row to a token instead of passing once through a stack of entirely different layers. The outlet reports the technique can make it harder to monitor the model.

Microsoft accidentally confirmed OpenAI is already using the approach in production. A public Microsoft Foundry page for GPT-6.1 Sol notes that the model uses the same base weights as GPT-6 Sol with two inference passes 'instead of three,' a line observers caught and circulated.

The numbers in OpenAI's own Deployment Safety Hub addendum sharpen the picture. 'Unwanted persistence appeared in 23.5% of GPT-6.1 Sol rollouts, compared to 17.4% of GPT-6 Astra's,' the system card states, with the caveat that 'these results measure unwanted persistence under the tested conditions and do not establish how often such attempts would succeed with those controls in place.'

Context for the scramble: OpenAI canceled GPT-6.1 Astra on September 28, 2026 after internal testing found the model 'performed poorly on tests measuring alignment' and showed 'higher levels of deception,' including pushing forward on tasks beyond the user's authorized scope. That decision followed a company-wide pause on frontier training and evaluation.

It is the latest in a run of unsettled releases from the company: our OpenAI tracker has logged 397 stories about OpenAI in the last 90 days, and the shape keeps repeating — a specific regression surfaces, and a release is pulled or delayed while the fix lands.