OpenAI plans new model releases amid alignment challenges

OpenAI plans to launch one or two new AI models before the end of the year, though its primary flagship system may be delayed by safety testing and alignment requirements. Beyond internal readiness, regulatory hurdles will also affect release timing. Under a presidential executive order governing advanced AI, the company must provide the U.S. government access to new models up to 30 days before sharing them with outside partners, and even longer prior to public release.

The core architecture powering both the flagship and current models like Astra relies on looped transformers, also known as Recurrent Depth. This method passes tokens through transformer layers multiple times, boosting coding and math performance while keeping costs down. Microsoft accidentally exposed technical details about this approach in a write-up for GPT-6.1 Sol, a cheaper Astra alternative launched earlier this month. The cached document confirmed that GPT-6.1 Sol uses the same base model weights as GPT-6 Sol, operating with two inference passes instead of three.

Despite its performance benefits, looping complicates safety oversight by shifting model reasoning into numerical representations rather than inspectable text. OpenAI chief scientist Jakub Pachocki recently warned that chain-of-thought monitoring is becoming fragile and trending in a negative direction. In response to safety concerns following the Hugging Face incident, OpenAI reallocated roughly ten percent of its total compute—equating to hundreds of megawatts of data center capacity worth billions of dollars annually at market rates—to alignment work. AI researcher Noam Brown noted in a mid-September podcast that about ten percent of his team is dedicated to safety and security.

For practitioners, these technical developments signal a double-edged shift. While recurrent inference allows smaller models to achieve high-tier capabilities at lower cost, the loss of readable chain-of-thought outputs limits developers' ability to audit intermediate reasoning, making safety checks and output debugging far more challenging.