OpenAI employees expect the company to ship “one or two new models” before the year is out, The Information reported, citing people familiar with the projects. Whether the major upgraded flagship it is building will be one of them is unclear. Here’s the practical upshot: the big model is still some way off, and smaller releases may arrive first.
Why the flagship is waiting
The new flagship needs alignment work so that its goals match what OpenAI’s researchers and customers expect, and the company must also strengthen the testing infrastructure around it. Even after leadership signs off on safety evaluations, another step remains. Under an executive order from President Donald Trump covering evaluations of advanced AI systems, OpenAI agreed to give the U.S. government access to new models up to 30 days before outside partners see them, and longer before any public launch.
Looping the same layers
The flagship, like Astra, the company’s current top model, uses looped transformers. OpenAI already runs the technique in production under the name “Recurrent Depth.” In plain terms, think of it like rereading a hard paragraph before answering. A conventional model sends each token through its layers once. A looped model sends it through the same layers several times before producing the next word. That lifts results in math and coding and lowers costs, because a smaller model can punch above its weight.
The trade-off is visibility. Looping moves part of the model’s thinking out of readable text and into internal numbers that human reviewers cannot inspect, which weakens chain-of-thought monitoring, the method safety researchers use to trace reasoning and catch bad behavior. The Information reports that OpenAI deliberately capped the number of loops in Astra to keep its reasoning readable. Chief scientist Jakub Pachocki has said this kind of monitoring is growing “fragile” and is “unfortunately trending in a negative direction.”
A leak from Microsoft
Details of how OpenAI uses looping surfaced through a Microsoft slip. The investor, which has access to OpenAI’s intellectual property, published a technical write-up on GPT-6.1 Sol, a cheaper alternative to Astra that OpenAI launched earlier this month. A section, since deleted, described “looping inference,” in which the model works the same problem through several passes before answering. According to a cached copy quoted by The Information, GPT-6.1 Sol uses the same base weights as GPT-6 Sol but runs two inference passes instead of three.
A person with direct knowledge confirmed the text matched what OpenAI had told Microsoft privately, and Microsoft pulled it after realizing it was not meant to be public. Looped transformers remain a sensitive area for OpenAI, in training and at inference.
Compute shifts toward safety
After the Hugging Face incident, OpenAI moved roughly ten percent of its total compute to alignment and safety, according to a person familiar with the figures. That amounts to hundreds of megawatts of data center capacity, though it is unclear what safety received before or how the split looks now. At market rates, renting that much compute would cost billions of dollars a year. Researcher Noam Brown said in a mid-September podcast that about ten percent of his team also works on security and safety.
Sources
- OpenAI will reportedly ship at least one model this year but its flagship still needs alignment work
