OpenAI reportedly plans one or two new models this year, but its flagship still needs alignment work

The Decoder, relaying reporting by The Information and citing people familiar with the projects, says OpenAI will likely release "one or two new models" before the end of the year, according to employees. It is unclear whether "the major upgraded model" the company is working on now will be among them.

The next flagship could still be a ways off. The model needs alignment work so its goals match what OpenAI's researchers and customers expect, and the company also has to shore up the testing infrastructure around it before it can move forward. Once leadership is satisfied with safety evaluations, there is another hurdle. 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 they go to outside partners, and longer still before a public launch.

The flagship also uses looped transformers, according to a person familiar with the situation. OpenAI already runs the technique in production under the name "Recurrent Depth", and it is central to how Astra, the company's current top model, works. Instead of passing a token through the transformer layers once, the model loops it through the same layers several times before outputting the next word. That raises performance on math and coding and lowers costs, because a smaller model can punch above its weight.

The cost is visibility. Looping moves some of the model's thinking out of readable text and into internal numerical representations that human reviewers cannot inspect, which undermines chain-of-thought monitoring, where safety researchers trace a model's reasoning to spot bad behavior. The Information reports that OpenAI deliberately capped how many loops Astra runs to keep its chain of thought readable.

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

The details of how OpenAI uses looping came out through a Microsoft slip-up. Microsoft, a major investor with access to OpenAI's intellectual property, published a technical write-up on GPT-6.1 Sol, a cheaper Astra alternative OpenAI launched earlier this month. A now-deleted section described GPT-6.1 Sol as using "looping inference", where the model works through the same problem in several passes before answering. Per a cached copy quoted by The Information, GPT-6.1 Sol "uses the same base model weights as GPT-6 Sol with two inference passes instead of three". A person with direct knowledge confirmed the write-up matched what OpenAI had told Microsoft in private, and Microsoft removed the section after realizing it was not meant to be public. Looped transformers remain a sensitive part of OpenAI's model development, both in training and at inference.

OpenAI chief scientist Jakub Pachocki has already acknowledged that chain-of-thought monitoring is growing "fragile" and "unfortunately trending in a negative direction". He also warned against an industry-wide race toward models without readable reasoning.

Key facts

OpenAI will likely release "one or two new models" before the end of the year, per employees cited by The Information; it is unclear whether the next major flagship is among them.
The flagship still needs alignment work and stronger testing infrastructure, and must clear a Trump executive order under which the US government gets access up to 30 days before outside partners.
The flagship reportedly uses looped transformers ("Recurrent Depth"), as Astra does; this helps math and coding and cuts costs but makes chain-of-thought monitoring harder.
Roughly ten percent of OpenAI's total compute, hundreds of megawatts of data center capacity, was reportedly moved toward alignment and safety after the Hugging Face incident.
Microsoft briefly published and then deleted a write-up saying GPT-6.1 Sol uses the same base weights as GPT-6 Sol with two inference passes instead of three.

Why it matters

The report ties OpenAI's release timing to safety work rather than to capability alone. The flagship's design choice, looping tokens through the same layers, makes models better at math and coding and cheaper to run, but it also pushes reasoning out of readable text and weakens chain-of-thought monitoring, one of the main ways safety researchers check what a model is doing. OpenAI's own chief scientist, Jakub Pachocki, has called that monitoring "fragile".

Who it affects

OpenAI customers and partners waiting on the next flagship are affected first, since the model is not ready and no date is given. The US government is in the loop under the executive order, getting access up to 30 days before outside partners. Safety researchers who rely on readable chains of thought are affected by the shift toward internal numerical reasoning, and Microsoft is involved as the investor whose deleted write-up exposed the looping details.

How to use it

There is nothing to use yet. No name or release date is given for the next flagship, and the report does not say which of the one or two models will ship. The model that already uses looping inference is GPT-6.1 Sol, the cheaper Astra alternative OpenAI launched earlier this month.

How solid is it

Low to moderate. Everything rests on unnamed employees and people familiar with the projects, relayed by The Decoder from The Information, and OpenAI and Microsoft comments are not given. One firmer point: a person with direct knowledge confirmed that the Microsoft write-up matched what OpenAI had told Microsoft in private, and the GPT-6.1 Sol wording comes from a cached copy of the deleted page.

Risks and caveats

The report does not say whether the flagship itself will ship this year, only that one or two models likely will. It is not stated how many loops Astra runs or what the cap is, and the article does not say what the Hugging Face incident was. It is also unclear how much compute went to safety before the shift or what the split looks like now. The Decoder frames the reduced readability of reasoning as a safety concern, while the compute figures come from a single person familiar with the numbers.

“unfortunately trending in a negative direction.”

— Jakub Pachocki, OpenAI chief scientist, on chain-of-thought monitoring, as relayed by The Decoder