OpenAI plans to ship one or two new AI models before the end of the year, though its upcoming flagship model remains delayed due to ongoing alignment and safety evaluations, according to internal sources cited by The Information. The delayed flagship utilizes a “looped transformer” technique—internally termed Recurrent Depth—which passes tokens through transformer layers multiple times to improve mathematical and coding capability while lowering compute costs.

The looping mechanism shifts part of the model’s intermediate reasoning into vector representations rather than human-readable text, hindering standard chain-of-thought monitoring and safety inspection. To address safety requirements, OpenAI reallocated roughly ten percent of its total compute capacity—equivalent to hundreds of megawatts—toward alignment research following recent security incidents.

Additionally, release schedules are subject to federal oversight under an executive order requiring OpenAI to grant the U.S. government up to 30 days of early model access prior to external partner deployment or public rollout.

Why it matters

  • Recurrent Depth and looped transformers trade chain-of-thought interpretability for improved reasoning efficiency and reduced operational costs.

  • Mandated 30-day federal safety review windows create explicit lag times between model completion and commercial availability.

  • Safety and alignment compute requirements now absorb significant portions of frontier lab data center capacity.

Source: the-decoder.com