OpenAI's GPT-6 Astra Debuts Looped Architecture and 'Hidden Reasoning'

OpenAI released GPT-6 Astra in September 2026, introducing a fundamental shift in how large language models process information. The model utilizes a "looped transformer" architecture that allows it to refine its understanding through internal iterations before producing a final answer.

Unlike traditional transformers that rely on a fixed number of layers, GPT-6 Astra employs recurrent depth. In this design, a shared block of layers is reused multiple times during inference, effectively turning the model's depth into a runtime variable. This allows the system to loop hidden states through the same layers repeatedly to refine its internal representation of a problem. This architectural change has resulted in improved multi-hop reasoning capabilities compared to its predecessor, GPT-5.6 Sol, even in instances where the model does not produce an explicit chain-of-thought output.

The Shift to Recurrent Depth

For years, the industry standard has been the vanilla transformer, which processes data through a linear stack of layers. While effective, these models have a rigid computational cost per token regardless of the problem's complexity. Looped transformers aim to solve this by allowing the model to "think" longer on harder problems by increasing the number of iterations through its shared layer block. This approach is designed to achieve a level of systematic generalization and efficiency that fixed-depth models typically lack.

The Challenge of Hidden Reasoning

This shift enables what researchers call "hidden reasoning." Because the model can refine its internal state through multiple passes without outputting a text-based trace, the reasoning process becomes a latent chain-of-thought. This creates a significant gap in interpretability; while a model using explicit chain-of-thought reveals its logic in the output, Astra's logic remains locked within its hidden states. By looping an entire transformer model on itself, the reasoning process is, by definition, hidden.

Implications for AI Safety

The inability to see the "trace" of how a model reached a conclusion raises urgent AI safety and auditing concerns. If the reasoning process is invisible, it becomes significantly harder for developers and regulators to monitor for hallucinations, biases, or deceptive logic. The lack of a visible audit trail means that the internal iterations act as a black box, complicating the effort to ensure the model is arriving at correct answers for the right reasons.

Performance Discrepancies

Early benchmarks have already highlighted tensions regarding the model's reported capabilities. Data from the ARC Prize indicates that GPT-6 Astra scored 62.7% on the standard provider-neutral harness. This stands in stark contrast to OpenAI's headline claim of 99.9%, which was achieved using a Provider Adapter harness—a difference of 37.2 percentage points. This discrepancy underscores the ongoing difficulty in establishing standardized, transparent metrics for models that utilize non-linear reasoning paths.