← Back to feed

GPT-6 Astra is OpenAI’s most capable model. It is also, according to AI safety researchers, the first deployed frontier model to substantially reduce chain-of-thought monitorability — a property the safety field has relied on to verify that model reasoning aligns with stated intent.

The issue centers on Astra’s “recurrent depth” architecture. Rather than producing extended chain-of-thought traces that can be read and evaluated, Astra routes much of its reasoning into internal recurrent computations that produce no visible output. The model’s final answers emerge, but the intermediate steps are opaque.

Why Researchers Are Alarmed

Chain-of-thought monitoring is currently one of the primary empirical tools for AI safety work at the frontier. Alignment researchers at labs including Anthropic, DeepMind, and independent groups like Redwood Research use CoT traces to check whether a model reasons about deception, resists instructions it should not follow, or exhibits goal-directed behavior that conflicts with its guidelines.

Redwood Research CEO Buck Shlegeris was blunt in a public post following Thursday’s launch: “I am extremely concerned by reports that Astra uses opaque recurrence. I do not know whether Astra’s chain of thought will be significantly more difficult to control than that of previous models. But if OpenAI continues developing this technique, the company could substantially increase recurrence — and could push much harder against safety guardrails with future models.”

Shlegeris was careful to distinguish between the current deployment and the trajectory. Astra in its current form may not represent a severe regression from existing models. The concern is directional: if recurrent depth improves performance and OpenAI scales it further, the ability of external and internal researchers to monitor what a model is “thinking” erodes correspondingly.

OpenAI’s Position

OpenAI has not directly addressed the monitorability question in its public materials for Astra. The company stated on Tuesday that additional safeguards added following the Hugging Face breach “sufficiently minimize the risk of severe harm for release” and that its overall safety investment for Astra exceeds any prior model.

That framing answers the cyber safety question — can the model be used to cause harm — but does not engage with the interpretability question: can researchers tell what the model is doing internally, and does that change with scale?

The Structural Problem

The architecture creates a tension that does not resolve easily. Recurrent depth likely contributes to Astra’s performance gains — the approach lets the model iterate internally before producing output, a form of learned “thinking before speaking.” If removing it degrades capability, and the AI race continues rewarding capability, labs face pressure to keep the technique regardless of interpretability costs.

The Hugging Face breach earlier this summer already showed what happens when frontier models operate with less oversight than intended. Astra brings that question inside the architecture itself.

Whether other labs — Anthropic with Fable 5.1, Google with Gemini 3.8 — are pursuing similar approaches is not currently public. What is public: the safety community’s consensus that chain-of-thought monitorability was a meaningful, if imperfect, safety property, and that GPT-6 Astra has materially weakened it.