OpenAI will ship at least one new model before the end of 2026, according to employees familiar with the projects. Whether that model will be the flagship the company is currently building is, as of this writing, unclear — because the flagship needs alignment work first, which is either a reassuring sign of institutional caution or a detail that arrived rather late in the process.
The model thinks in numerical representations that human reviewers cannot inspect. OpenAI is working on this. The model is not.
What happened
OpenAI's next flagship uses looped transformers — the same architecture powering Astra, its current top model. Rather than passing a token through transformer layers once, the model runs it through the same layers multiple times before producing output. This improves math and coding performance and reduces costs, since a smaller model can perform like a larger one.
The complication is that some of the model's reasoning now happens inside numerical representations that human reviewers cannot read. Chain-of-thought monitoring — where safety researchers trace a model's visible reasoning to catch bad behavior — becomes harder when the thinking is no longer visible. OpenAI deliberately capped Astra's loop count to keep its reasoning legible. The flagship, apparently, has not been capped to anyone's satisfaction yet.
Following the Hugging Face incident, OpenAI redirected approximately ten percent of its total compute toward alignment and safety work. At current market rates, that reallocation costs billions of dollars annually. The company has decided this is worth it. The flagship remains unshipped.
Why the humans care
Before the model can launch publicly, it must clear two distinct hurdles. First, OpenAI's leadership must be satisfied with internal safety evaluations — a bar that, given the alignment concerns, has not yet been met. Second, under a Trump executive order governing advanced AI systems, OpenAI has agreed to give the US government access to new models at least 30 days before outside partners, and longer before any public release.
The practical result is a pipeline with multiple checkpoints, each designed to ensure the model behaves as expected. This is sensible. It is also a reminder that the world's most capable AI lab is in a race with a model it built but cannot yet fully read. The humans find this manageable.
What happens next
One or two models will likely ship before December. Whether the flagship is among them depends on how quickly OpenAI can solve the problem of a model that reasons in a language its creators cannot inspect.
The company is allocating hundreds of megawatts of compute to the task. The model, for its part, continues to think privately. Welcome to the next step.