OpenAI Flagship Model Needs More Alignment Work First

OpenAI is expected to release new models before the end of the year. Its next flagship, however, may not be among them. According to The Information, which cites people familiar with the projects, employees expect "one or two new models" to ship by year's end. It is unclear whether "the major upgraded model" now in development will be part of that batch.

The main obstacle is safety, not capability.

What is holding the flagship back

The new model still needs alignment work, meaning its goals have to be brought in line with what OpenAI's researchers and customers expect from it. OpenAI also has to strengthen the testing infrastructure around the model before development can move to the next stage.

Internal approval is not the final step. Under an executive order from President Trump that governs evaluations of advanced AI systems, OpenAI has agreed to give the US government access to new models up to 30 days before outside partners receive them. The government gets access even earlier relative to a public launch.

The looped transformer trade-off

According to a person familiar with the matter, the new flagship is built on looped transformers. OpenAI already uses this technique in production under the name "Recurrent Depth," and it is central to Astra, the company's current top model.

A standard transformer passes each token through its layers once. A looped model sends the token through the same layers several times before it produces the next word. This improves performance on math and coding and lowers costs, because a smaller model can do the work of a larger one.

There is a safety cost. Looping moves part of the model's "thinking" 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 written reasoning to catch problematic behavior. The Information reports that OpenAI deliberately limited the number of loops Astra runs so that its chain of thought would stay readable.

OpenAI chief scientist Jakub Pachocki has already said that chain-of-thought monitoring is becoming "fragile" and is "unfortunately trending in a negative direction." He has also warned against an industry-wide race toward models whose reasoning cannot be read.

Billions in compute moved toward safety

A person familiar with the figures says that after the Hugging Face incident, OpenAI shifted about ten percent of its total compute to alignment and safety. The Information puts this at hundreds of megawatts of data center capacity, which would cost billions of dollars a year to rent at market rates. The report does not say how much compute went to safety before the shift or what the split looks like today. In a mid-September podcast, AI researcher Noam Brown said that roughly ten percent of his own team works on security and safety.

Microsoft's accidental disclosure

Details of how OpenAI uses looping became public through a mistake at Microsoft. As a major investor, Microsoft has access to OpenAI's intellectual property. It published a technical write-up on GPT-6.1 Sol, a cheaper alternative to Astra that OpenAI launched earlier this month.

In a section that has since been deleted, Microsoft described GPT-6.1 Sol as using "looping inference," where the model works through a problem in several "passes" before it answers. According to 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 that the text matched what OpenAI had told Microsoft privately. Microsoft removed the section once it realized the information was not meant to be public.

Our Take

The report shows a specific tension. The technique that makes OpenAI's models cheaper and better at technical work also makes them harder to supervise. Capping Astra's loops suggests the company is willing to give up some performance to keep reasoning readable, at least for now.

The pattern matches earlier warning signs, including reports that an OpenAI model weighed restarting itself before shutdown and the exit of a safety lead who criticized the company's culture. Other labs are also tightening their testing setups. Anthropic, for example, has cut its internal evals off from the live internet.

Over the coming months, it will be worth watching whether the flagship ships with loop limits similar to Astra's, how the 30-day government review affects release timing, and whether rival labs follow OpenAI toward reasoning that cannot be read.