OpenAI Plans to Ship One or Two Models This Year as Flagship Needs Alignment Work

Model Timeline

OpenAI plans to release "one or two new models" before the end of the year, according to employees cited by The Information. It is unclear whether the company's major upgraded flagship will be among them. The flagship still needs alignment work so its goals match what OpenAI's researchers and customers expect, and the company must also shore up testing infrastructure around the model.

Under an executive order from President Trump governing evaluations of advanced AI systems, OpenAI agreed to give the US government access to new models up to 30 days before outside partners, and longer still before a public launch.

Looped Transformers and Safety

The flagship uses the same looped transformer technique as Astra, OpenAI's current top model, where the method is called "Recurrent Depth." Instead of passing a token through the transformer layers once, the model loops it through the same layers multiple times before outputting the next word. The approach boosts math and coding performance and lowers costs, since a smaller model can punch above its weight.

The tradeoff is oversight. Looping moves some of the model's "thinking" out of readable text and into internal numerical representations that human reviewers cannot inspect, undermining chain-of-thought monitoring. OpenAI deliberately capped how many loops Astra runs to keep its chain of thought readable. Chief scientist Jakub Pachocki has acknowledged that chain-of-thought monitoring is growing "fragile" and "unfortunately trending in a negative direction," and he warned against an industry-wide race toward models without readable reasoning.

Compute Shift and Microsoft Leak

After the Hugging Face incident, OpenAI shifted roughly ten percent of its total compute toward alignment and safety, a person familiar with the figures said. That amounts to hundreds of megawatts of data center capacity, though it is unclear how much was allocated to safety before or what the current split looks like. At market rates, the shifted compute would cost billions of dollars a year to rent. AI researcher Noam Brown said in a mid-September podcast that about ten percent of his team also works on security and safety.

Microsoft, a major investor with access to OpenAI's intellectual property, inadvertently revealed how the technique works. In a now deleted section of a technical write up on GPT-6.1 Sol, a cheaper Astra alternative launched earlier this month, Microsoft described "looping inference," where the model works through the same problem in multiple "passes." The page said GPT-6.1 Sol "uses the same base model weights as GPT-6 Sol with two inference passes instead of three." Microsoft removed the section after realizing it was not meant to be public.

Sources: The Decoder