OpenAI's Next Flagship Is Stuck on Alignment, Report Says

OpenAI will ship at least one new model by the end of the year, even if its next flagship model is not ready, employees told The Decoder. The flagship, which is still in development, uses the same looped transformer technique as Astra to boost math and coding performance.

The technique carries a documented tradeoff. Looping makes safety monitoring harder, according to the report, which is why the flagship still needs alignment work (the process of ensuring a model behaves as intended) before release.

Microsoft complicated matters. The company accidentally revealed details about how the looping technique works on a public page, exposing internal mechanics of an unreleased system.

What do we know about the release plan?

• One or two new OpenAI models will likely ship by year's end, according to employees.
• The flagship model's readiness remains uncertain; it depends on completing alignment work.
• The model targets improved math and coding performance through looped transformers.

Why does looping complicate safety monitoring?

A looped transformer reuses computation across repeated passes rather than running a single fixed forward pass. That iterative structure, The Decoder reports, makes it harder for OpenAI to monitor what the model is doing internally — a direct obstacle to alignment (safety verification) work.

The stakes are concrete. OpenAI's flagship models anchor its commercial roadmap, and a technique that raises performance while degrading monitorability forces a tradeoff between capability shipping timelines and safety assurance.

The reported plan suggests OpenAI intends to keep its release cadence on schedule with other models, regardless of when the flagship clears alignment review.