Report: OpenAI Hardens Safety for Unreleased Model Using Loop Transformer

OpenAI is working on alignment and security hardening to safely release an unreleased next-generation AI model, according to a report by The Information on October 13. The report says the model uses a Loop Transformer design, also called Recurrent Depth, which reuses the same neural network layers repeatedly.

The reported architecture may hide internal computations from the human-readable chain-of-thought, potentially weakening existing monitoring for anomalous behavior. OpenAI has not officially confirmed using the Loop Transformer technique. It reportedly limited the use of recurrent structures in its Astra model to keep human-readable reasoning records.

Microsoft reportedly described in a blog post last week that GPT-6.1 Sol uses the same base model weights as GPT-6 Sol but passes through reasoning twice instead of three times, then later deleted that description.

Safety work and compute

OpenAI reportedly allocated about 10 percent of its total computing resources to alignment and safety work after the Hugging Face attack incident late last month, equivalent to hundreds of megawatts of data center capacity. The company is also strengthening sandboxes that isolate unreleased models from the internet and other software, and has changed its model training start procedure so that responsible people document risks and reasons for safe training with multiple sign-offs. Each responsible person can independently refuse training, and systems to automatically halt training or notify employees upon incidents are already in place or planned, according to the report.

An employee said that if the next-generation model passes internal safety evaluations, it could be provided to the US government for up to 30 days of testing, then shared with trusted partners and later widely distributed. Employees said one or two new models are likely to be released by the end of the year, though they may not be major upgrade models.

Separately, Google engineer Patrick Toomey recently said on X that developers at a New York AI industry gathering expect Anthropic's Fable 5.5 to show great performance but OpenAI's Bel, to be unveiled in December, will surpass it. Bel has been a codename known since August, when posts on X claimed OpenAI had completed pre-training of Bel with over 10 trillion parameters for use as a foundation for future AGI-level models, and some media reported this. OpenAI CEO Altman said in a Time interview around the same time that the company would build a system it could internally call AGI by the end of the year. Bel's specifications and release schedule have not been officially confirmed, and it is unconfirmed whether the next-generation model reported by The Information is Bel.