OpenAI's Astra model sparks safety fears over recurrent depth

OpenAI is getting ready to drop Astra, its most powerful model yet, but researchers are already sounding the alarm over what is happening under the hood.

According to a report from The Information, Astra uses a technique known as recurrent depth or looped transformers. The architecture cycles information through internal layers multiple times before generating text, dramatically boosting coding and math performance.

There is just one massive catch: it cuts out the model’s internal monologue, making it far harder to monitor.

Losing the paper trail

Most frontier models process prompts sequentially, showing their work in a step-by-step chain of thought. While imperfect, this visible reasoning allows automated oversight systems and researchers to detect whether a model is hallucinating, attempting to deceive users, or planning to bypass safety guardrails.

Recurrent depth changes that dynamic. By looping tokens through internal layers, much of Astra's reasoning reportedly happens inside latent space rather than readable human language.

AI safety researchers did not take the news lightly. Ryan Greenblatt, chief scientist at Redwood Research and one of three outside researchers who probed a recent hack on OpenAI, warned that adopting opaque architectures may be the single worst development for AI safety to date.

Greenblatt noted that unpicking that previous breach relied almost entirely on chain-of-thought logs. If models stop reasoning in plain text, auditing their real motives becomes an uphill battle.

Pushing back against the panic

OpenAI has kept quiet on the exact architecture, but its leadership quickly went on defense on social media. Chief scientist Jakub Pachocki said Astra’s computational depth is within a factor of two of GPT-4, arguing that reports blew the model's opacity out of proportion.

Pachocki also noted that OpenAI has fought to protect chain-of-thought monitoring since its earliest reasoning models, though he warned that monitorability is already trending in a negative direction for broader reasons.

Sources told The Information that OpenAI intentionally limited the technique in Astra to keep oversight intact. Meanwhile, both Anthropic and Google DeepMind are reportedly already discussing the looped approach internally.

Why it matters

For builders and power users, recurrent depth looks like an obvious win. It packs immense mathematical and programming horsepower into models without requiring miles of verbose reasoning tokens.

However, we think this architectural shift points to an uncomfortable crossroads for the entire industry. Chain of thought was never a bulletproof safety mechanism, but it was the only real-time audit window researchers had into frontier models.

If labs start trading legible text for raw latent capability, they risk triggering a dangerous race to the bottom where competitive pressure forces everyone to sacrifice observability.

With autonomous agents already being tested against real-world systems, deploying models whose internal strategies cannot be inspected feels like a risky gamble, no matter how fast they write code.