OpenAI may release one or two models before the end of the year, according to employees cited by The Information. Whether the company’s major upgraded model will be among them remains unclear: it still needs alignment work and stronger testing infrastructure.
Release timing depends on safety work
The next flagship model is not ready to move forward simply because it is being developed. Its goals need to match what OpenAI’s researchers and customers expect, and the company must strengthen the testing systems around it.
Those steps are part of the release process described in the report. Even after leadership is satisfied with safety evaluations, the model faces another review stage: 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.
That leaves room for other models to arrive first. The reported expectation of one or two releases by year’s end does not establish that the flagship itself will be ready on that schedule.
Repeated passes can boost performance
The flagship is said to use looped transformers, a technique OpenAI already runs in production under the name “Recurrent Depth.” It is central to how Astra works, according to the report.
In a conventional pass through transformer layers, a token moves through the layers once before the model produces the next word. With looping, it passes through the same layers multiple times. The report says this can improve math and coding performance while lowering costs, because a smaller model can deliver more capability.
Microsoft’s now-deleted technical write-up offered a glimpse of how the method is used. It described GPT-6.1 Sol as using “looping inference,” working through the same problem in multiple “passes.” The Information quoted a cached copy 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 information OpenAI had shared privately with Microsoft. Microsoft removed the section after realizing it was not meant to be public. The report says looped transformers remain sensitive in OpenAI’s work on both training and inference.
More computation can mean less readable reasoning
The same repeated processing that may help a model solve problems creates a monitoring challenge. Some of its “thinking” takes place in internal numerical representations rather than readable text, making it harder for human reviewers to inspect.
That matters for chain-of-thought monitoring, in which safety researchers trace a model’s reasoning to look for bad behavior. If more reasoning is hidden from view, reviewers have less readable material to assess. The Information reports that OpenAI deliberately capped the number of loops Astra runs to keep its chain of thought readable.
OpenAI chief scientist Jakub Pachocki has acknowledged that chain-of-thought monitoring is growing “fragile” and “unfortunately trending in a negative direction.” He has also warned against an industry-wide race toward models without readable reasoning. His comments point to a tension in the design: improving capability through additional passes can make one established way of checking model behavior less dependable.
Safety resources are part of the effort
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 described that allocation as hundreds of megawatts worth of data center capacity.
The report says the earlier amount allocated to safety, and the current split, are unclear. It also says renting the shifted compute at market rates would cost billions of dollars a year. Those figures give a sense of the scale described, while leaving the precise before-and-after allocation unknown.
AI researcher Noam Brown said in a mid-September podcast that about ten percent of his team also works on security and safety. The report does not say that this team share is the same measure as OpenAI’s compute allocation.
The company’s next release therefore hinges on more than model performance. OpenAI is working with a technique that may improve math and coding results and reduce costs, while requiring stronger safety evaluations and posing challenges for readable-reasoning checks. The reported release window applies to one or two new models; the flagship’s timing remains uncertain.