OpenAI’s Astra’s “innovation” was actually already publicly demonstrated by ByteDance last year. This approach is called recurrent depth. Simply put, instead of relying solely on generating longer chains of thought (CoT) to think longer, it repeatedly cycles the same set of Transformer blocks over the hidden state, packing more computation into the latent space. Harder problems run more rounds; easier ones exit early. Parameters don’t need to scale with reasoning depth, and test-time compute can be allocated more flexibly. In October last year, ByteDance’s Seed publicly released and open-sourced Ouro—a looped language model. Ouro-1.4B and 2.6B incorporated iterative latent reasoning directly during pretraining; the 2.6B model defaults to multiple rounds of recurrent computation while supporting adaptive exit, allowing each problem to determine how many rounds it needs. Last year, this was merely a side project validated by ByteDance using smaller models; now, OpenAI has integrated it into its most advanced model. This suggests that recurrent depth is no longer just an academically “interesting” design—it’s beginning to enter the true engineering roadmap for the next generation of large models. But the challenges it brings are more complex than benchmark results suggest. Previously, scaling reasoning models often involved embedding computations directly into CoT. The longer the model thought, the more reasoning tokens it produced. Although CoT isn’t equivalent to the model’s full internal state, it at least provided a visible monitoring window. Recurrent depth continues shifting computation deeper into the latent space. The same parameters can cycle many times within the hidden state, completing substantial internal computation before outputting only a short text segment. This raises a very practical question: The actual location of the model’s reasoning is gradually moving away from natural language. This is precisely what researchers now fear most about Astra. OpenAI has long emphasized Chain-of-Thought monitoring because when models intend to cheat, hack rewards, or circumvent rules, they often first reveal those intentions in their CoT. But if more and more reasoning occurs within latent states, what exactly can you monitor? Reports suggest OpenAI is currently deliberately limiting the extent to which Astra uses recurrent depth—to preserve more readable CoT and to deploy additional monitoring of internal states and behaviors. This is actually quite fascinating. For years, scaling reasoning in large models was straightforward: give them more tokens so they can say more. The next phase may involve giving them more internal cycles—without requiring them to tell you anything about the process.