OpenAI will likely release one or two new models before the end of the year, though its next flagship model still requires significant alignment work. According to employees and reports from The Information, the upcoming flagship utilizes a looped transformer technique known internally as Recurrent Depth, which is also used in the Astra model. This method loops tokens through the same transformer layers multiple times before generating output, significantly boosting math and coding capabilities while lowering operating costs.
However, this looped architecture complicates safety monitoring by moving part of the model's reasoning into internal numerical representations that human reviewers cannot easily inspect. This undermines traditional chain-of-thought monitoring, prompting OpenAI to cap the number of loops on current models to maintain readability. Following a security incident involving Hugging Face, OpenAI dedicated roughly ten percent of its total compute capacity, amounting to hundreds of megawatts, toward alignment and safety efforts. Additionally, the company must provide the US government access to new models up to thirty days prior to external partner releases, complying with an executive order governing advanced artificial intelligence systems.
The inner workings of this technique were recently highlighted when investor Microsoft accidentally published and later deleted details regarding GPT-6.1 Sol, confirming its use of looping inference across multiple passes. OpenAI chief scientist Jakub Pachocki has publicly acknowledged that monitoring model chain-of-thought reasoning is becoming increasingly fragile. Why it matters: Balancing advanced reasoning performance with transparent safety alignment remains a critical technical hurdle for top artificial intel