OpenAI is reportedly planning to release at least one model this year, but its flagship model still requires alignment work.

Meta description: OpenAI may ship one or two new models before year's end as its next flagship uses looped transformers, raising capability gains and safety monitoring concerns.

Tags: OpenAI, looped transformers, AI safety, Astra, GPT-6.1 Sol

Featured image: An abstract layered neural network looping through transformer blocks, with soft blue and amber gradients, no text, no logos, no people.

OpenAI is working to complete its next flagship model, and employees say the company will likely release one or two new models before the end of the year. The plans matter because the flagship is built around the same looped transformer approach used by Astra, OpenAI's current top model, a design that can strengthen math and coding performance while making safety monitoring harder.

Background: A flagship built on the same technique as Astra

The Information, citing people familiar with the projects, reported that OpenAI expects one or two new models by year's end. It remains unclear whether the company's major upgraded model will be among them. The new flagship could still be some distance from release. It needs alignment work so its goals match what OpenAI's researchers and customers expect. OpenAI also needs to strengthen the testing infrastructure around the model before it can move forward.

The next flagship uses looped transformers, according to a person familiar with the situation. OpenAI already runs this technique in production under the name Recurrent Depth. The method is central to Astra. 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. That can raise performance on math and coding and reduce costs because a smaller model can achieve more than its size would suggest.

Development: Safety and government review shape the release path

The same looping creates a safety problem. It moves parts of the model's reasoning out of readable text and into internal numerical representations that human reviewers cannot inspect. That weakens chain-of-thought monitoring, a method in which safety researchers follow a model's reasoning to detect harmful behavior. The Information reports that OpenAI deliberately limited the number of loops Astra runs to keep its chain of thought readable.

The next hurdle is government review. Once leadership is satisfied with safety evaluations, an executive order from President Trump governing evaluations of advanced AI systems requires another step. Under that order, OpenAI agreed to give the US government access to new models up to 30 days before they go to outside partners, and longer before a public launch.

The company's safety priorities also changed after the Hugging Face incident. A person familiar with the figures said OpenAI moved roughly ten percent of its total compute toward alignment and safety. The Information reports that this equals hundreds of megawatts worth of data center capacity, though it is unclear how much compute 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.

Latest progress: A Microsoft leak reveals looped inference details

OpenAI's internal use of the technique became public through an error at Microsoft. 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.

In a now-deleted section, Microsoft described GPT-6.1 Sol as using looping inference, where the model works through the same problem in multiple passes before producing an answer. The Information quoted a cached copy of the page saying 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 that the write-up matched what OpenAI had told Microsoft privately. Microsoft removed the section after realizing it was not meant to be public.

The leak showed that looped transformers remain a sensitive part of OpenAI's model development, both during training and at inference. OpenAI chief scientist Jakub Pachocki has acknowledged that chain-of-thought monitoring is growing fragile and is unfortunately trending in a negative direction. He also warned against an industry-wide race toward models without readable reasoning.

Where things stand

OpenAI is at a pre-release stage. Employees point to one or two new models likely shipping by year's end, but the major upgraded flagship still needs alignment, stronger testing, safety approval, and government review access. The looped transformer approach gives OpenAI a route to stronger math and coding results at lower cost, while also pushing safety monitoring into less visible internal representations. The next question is whether the company can ship before those trade-offs are resolved.