OpenAI Prepares New Releases as Looped Transformers Complicate Flagship Alignment

Standard transformer models pass tokens through network layers a single time. By contrast, a looped transformer feeds tokens through the identical layers multiple times before generating the next word. This iterative processing significantly enhances performance on rigorous tasks such as mathematics and software development while reducing operational costs, allowing smaller models to match the output of larger networks.

However, the recursive technique creates distinct safety and oversight challenges. When a model loops through internal layers, aspects of its reasoning migrate away from readable natural language into raw numerical states that human evaluators cannot directly inspect. This shift weakens chain-of-thought monitoring, an essential safety technique researchers use to audit the step-by-step logic of an AI system. To prevent its chain of thought from becoming completely unreadable, OpenAI intentionally restricted the number of recurrent loops permitted in Astra.

OpenAI chief scientist Jakub Pachocki has publicly pointed out these architectural risks, noting that chain-of-thought visibility is becoming increasingly "fragile" and "unfortunately trending in a negative direction." Pachocki has cautioned against an industry-wide push toward deploying high-performing models that lack transparent reasoning trails.

Safety compute and the Microsoft technical leak

To address safety concerns, OpenAI reallocated approximately ten percent of its total computing power toward alignment and safety initiatives following an incident involving Hugging Face. That computing shift represents hundreds of megawatts of data center infrastructure, which would cost billions of dollars annually at current market hosting rates. Furthermore, AI researcher Noam Brown noted in a mid-September podcast that roughly ten percent of his team focuses specifically on safety and security.

Technical verification of how OpenAI deploys recurrent inference emerged publicly through an inadvertent leak by Microsoft. As an investor with access to OpenAI's intellectual property, Microsoft published a technical overview detailing GPT-6.1 Sol—a lower-cost alternative to Astra released earlier this month.

Before the section was deleted, Microsoft disclosed that GPT-6.1 Sol utilizes "looping inference," resolving queries across several iterative passes. According to cached documentation, GPT-6.1 Sol relies on the identical base model weights as GPT-6 Sol, but completes two inference passes rather than three. Sources confirmed that the removed material accurately reflected private technical briefings provided by OpenAI.

What it means for developers

For developers building applications, OpenAI's emphasis on looped inference signals a technical shift where models can deliver higher coding and math capabilities at lower costs. A model with fewer base parameters running multiple passes can operate more efficiently than a substantially larger dense network.

At the same time, this technique introduces new trade-offs in explainability. Because internal looping suppresses explicit chain-of-thought tokens in favor of latent numerical representations, teams relying on intermediate reasoning for validation or auditability may find inspection harder to maintain.

As model architectures evolve and new options arrive, developers can try top AI models cheaply through one API at https://apixoai.online to evaluate performance, reasoning clarity, and cost efficiency across different providers without committing to a single infrastructure provider.

Source: OpenAI will reportedly ship at least one model this year but its flagship still needs alignment work — The Decoder. Written by the Apixo team from that report.