OpenAI is likely to release one or two new models this year, according to people at the company. It remains unclear, however, whether the flagship model in development will be among them. The flagship is said to still need work to align its behavior with the expectations of researchers and customers, as well as upgrades to its testing infrastructure. The report was published on October 14, 2026.
The Models Due This Year—and the Flagship
The key point to take away is that “a new model is coming” and “the flagship model is coming” are two different things. Employees believe one or two models are likely to be released by the end of the year. But it is still uncertain whether the major upgrade currently in development will be among them.
The flagship has two major hurdles left. One is alignment: making the AI’s objectives match what its developers and business customers want. The other is strengthening the testing environment used to evaluate the model. Neither shows up in performance benchmarks, but both are steps that can determine whether a model ships.
There is another hurdle even after executives are satisfied with the company’s internal safety evaluations. Under a framework for evaluating advanced AI established under an executive order by President Trump, OpenAI has reportedly agreed to give the U.S. government access to new models up to 30 days before sharing them with external partners. A longer wait should be expected before public release. This is no longer an era when a model can be “released as soon as it’s finished.”
A System That Repeats the Same Process to Refine Its Answers
A key technology behind the flagship model is the “looping transformer.” Conventional models pass tokens—the smallest units of language—through a stack of computational layers just once to generate the next word. A looping transformer passes them through the same layers repeatedly before producing the next word. OpenAI is already using the approach in production under the name “Recurrent Depth,” and it is reportedly at the core of its current top-tier model, Astra.
The goal is to improve performance in math and coding while reducing costs. Even a small model can deliver more advanced capabilities by thinking repeatedly. It is somewhat like reading the same textbook several times to deepen your understanding.
The Better the Performance, the Less Transparent the Model Becomes
But this approach comes at a cost. Some of the model’s “thinking” takes place not in human-readable text, but in internal numerical representations. Safety researchers have traditionally tracked the written reasoning a model produces on its way to an answer to identify suspicious behavior. This is known as monitoring the “chain of thought.” The more loops a model runs, the harder it becomes to monitor.
OpenAI reportedly capped the number of loops in Astra to keep its reasoning readable. In effect, it shipped the model while deliberately leaving room to improve its performance. The flagship is also said to use the same technology, making where to draw the line the next key question.
What Microsoft’s Deleted Documentation Revealed About Confidentiality
Microsoft technical documentation brought the technology into public view. Microsoft, which has access to OpenAI’s intellectual property, published an explanation of “GPT-6.1 Sol,” a lower-cost version of Astra that OpenAI released this month. The documentation discussed “looped reasoning,” explaining that the model processes the same problem through multiple “passes” before producing an answer.
According to the description, GPT-6.1 Sol used the same base-model weights as GPT-6 Sol but reduced the number of reasoning passes from three to two. That section was later removed. A person familiar with the matter confirmed that the description matched information OpenAI had shared confidentially. The episode suggests that looping technology is an area OpenAI handles cautiously, both in training and inference.
Varying the number of passes to create products at different price and performance levels is a practical way to build multiple offerings from the same foundation. At the same time, the more passes the model makes, the less transparent its inner workings become. Product design and safety design are connected to the same dial.
The Figure of 10% of Computing Resources
Specific figures have also been reported on investment in safety. After an incident involving Hugging Face, OpenAI is said to have allocated about 10% of its total computing resources to alignment and safety. In terms of data-center power, that amounts to several hundred megawatts, which would cost billions of dollars a year at market rates.
Researcher Noam Brown also said on a podcast in mid-September that about 10% of his team works on security and safety. The figure of 10% is becoming a benchmark for the scale of the organization’s efforts.
Can Readable Reasoning Be Preserved?
Chief Scientist Jakub Pachocki has acknowledged that monitoring the chain of thought is becoming “fragile” and is “unfortunately moving in the wrong direction.” He has also warned against a race in which the entire industry pushes toward models that lack readable reasoning.
Improved performance and lower costs are attractive from a business perspective. Precisely because that appeal is so strong, the question is how far the models can advance without giving up monitorability. The one or two models expected to arrive this year should offer a progress report of sorts. I’m more interested in what condition the flagship model is released in than in when it comes out.
Comments