SAN FRANCISCO, October 15, 2026 — OpenAI plans to release one or two new AI models by the end of the year, though its next flagship system remains delayed due to ongoing alignment and safety work.
New Models on the Horizon
Employees at OpenAI say the company will ship at least one model before year’s end, with the possibility of a second following shortly after. These releases are separate from the larger flagship model currently in development, which sources say is still being refined. The upcoming models are expected to build on existing technology while addressing immediate needs from partners and customers.
Flagship Model Faces Alignment Hurdles
The next flagship model requires additional alignment work to ensure its behavior matches OpenAI’s research goals and user expectations. This process involves adjusting the model’s outputs so they remain safe, reliable, and aligned with intended use cases. Until this work is complete, the flagship cannot proceed to wider testing or release.
Looped Transformers Power Performance Gains
Both the upcoming flagship and OpenAI’s current top model, Astra, use a technique called looped transformers, internally referred to as “Recurrent Depth.” Instead of processing each token once, the model loops it through the same transformer layers multiple times before generating the next word. This approach improves performance on tasks like math and coding while allowing a smaller model to achieve stronger results, reducing computational costs.
Safety Monitoring Becomes More Complex
However, the looping technique moves some of the model’s reasoning into internal numerical states that are not visible in text, making it harder for safety teams to monitor its chain of thought. OpenAI chief scientist Jakub Pachocki has acknowledged that this form of monitoring is becoming “fragile” and is “unfortunately trending in a negative direction.” To mitigate risks, OpenAI has deliberately limited the number of loops in Astra to keep reasoning more interpretable.
Government Access and Internal Shifts
Under an executive order from President Trump, OpenAI has agreed to provide the U.S. government with up to 30 days of pre-release access to new models before they are shared with external partners. Following the Hugging Face incident, the company shifted approximately ten percent of its total compute resources toward alignment and safety efforts. This reallocation reflects a broader focus on strengthening safeguards as model capabilities advance.
Technical Details of Looped Inference Revealed
Microsoft accidentally disclosed details about OpenAI’s looped inference technique through a now-deleted technical write-up on GPT-6.1 Sol, a cheaper Astra alternative launched earlier in the month. The document described GPT-6.1 Sol as using “looping inference,” where the model works through the same problem in multiple “passes” before producing an answer. Specifically, it noted that GPT-6.1 Sol uses the same base model weights as GPT-6 Sol but with two inference passes instead of three. A person with direct knowledge confirmed the write-up matched what OpenAI had shared privately with Microsoft, and the section was removed after Microsoft realized it was not intended for public release. Looped transformers remain a sensitive aspect of OpenAI’s model development, both during training and at inference.
Compute Reallocation and Cost Implications
Following the Hugging Face incident, OpenAI moved roughly ten percent of its total compute resources toward alignment and safety efforts. This shift amounts to hundreds of megawatts of data center capacity, though the exact allocation before and after the change is unclear. At market rates, the compute resources reallocated would cost billions of dollars annually to rent. AI researcher Noam Brown noted in a mid-September podcast that about ten percent of his team also works on security and safety, reflecting a broader industry trend toward prioritizing safeguards as model capabilities grow.
Sources
- OpenAI will reportedly ship at least one model this year but its flagship still needs alignment work
Photo: TechCrunch (CC BY 2.0)
