OpenAI may ship a model this year as its flagship remains in testing

The cost of making reasoning legible

The reported flagship uses recurrent transformers, a technique OpenAI already employs in Astra under the name “recurrent depth.” Instead of passing a token through each transformer layer once, the model sends it through the same layers multiple times before producing the next word.

That approach can improve performance in math and coding while reducing costs: a smaller model can take on work that would usually require a larger one. But some of its “thinking” moves from readable text into internal numerical representations that people cannot inspect. That makes it harder for safety researchers to trace the model’s reasoning and spot unwanted behavior. The Information reports that OpenAI deliberately limited the number of cycles in Astra to keep its reasoning chain readable.

The trade-off is not merely technical. More cycles may help a model perform, but can make its behavior harder to audit. OpenAI chief scientist Jakub Pachocki has acknowledged that control over reasoning chains is becoming “fragile” and, unfortunately, getting worse. He has also warned against an industry-wide race toward models whose reasoning cannot be read.

After an incident involving Hugging Face, OpenAI directed about 10% of its computing resources to alignment and safety, a person familiar with the figures told The Information. The outlet says that is equivalent to hundreds of megawatts of data-center capacity. The company’s previous allocation and its current distribution of resources are not known.

At market rates, renting that redirected capacity would cost billions of dollars a year. AI researcher Noam Brown said in a podcast in mid-September that about 10% of his team also works on alignment and safety. Those figures suggest the work has a substantial resource footprint, but they do not show how much OpenAI’s safety effort has changed over time.

A disclosure that exposed the method

More detail about OpenAI’s use of recurrent inference surfaced after Microsoft, a major OpenAI investor with access to its intellectual property, published a technical description of GPT-6.1 Sol. OpenAI released the model earlier this month as a lower-cost alternative to Astra.

Microsoft’s subsequently removed description called GPT-6.1 Sol a model with “recurrent inference,” breaking a task into multiple passes before answering. A saved copy said it uses the same base-model weights as GPT-6 Sol but makes two inference passes instead of three, The Information reports.

A person directly familiar with the details confirmed that the description matched information OpenAI had shared privately with Microsoft. Microsoft removed the section after realizing it was not meant for publication. The Information reports that recurrent transformers remain sensitive to OpenAI’s model development, both in training and inference.

I think the timing puts the central tension in sharper focus: OpenAI is trying to get more capability from repeated computation while keeping the reasoning legible enough to inspect. The announcement is quiet about how the company will measure that balance in its flagship, or what level of visibility safety reviewers will retain as the model’s cycles increase. Until those details emerge, the promise of cheaper, stronger reasoning comes with an unresolved cost: less of the process may be available for people to examine.