Reported OpenAI training technique that reuses the same transformer layers repeatedly, reasoning in latent space.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Recurrent depth, also called a looped transformer, is an architecture and training approach in which a model passes its activations through the same block of layers multiple times instead of using only a fixed stack of distinct layers. This lets the model spend more computation per token at inference time, with reasoning happening in its hidden state rather than in written-out text.
According to a report from The Information, cited by a YouTube creator (Matt Wolfe), OpenAI reportedly used a technique of this kind for a model referred to as "Astra", which processes the same text multiple times. The creator said this could obscure chain-of-thought reasoning, which raises concerns about monitoring and safety, since reasoning that is not expressed in readable tokens is harder to inspect. These details come from secondary reporting and have not been confirmed through an official OpenAI page.
No official documentation, paper, code or weights from OpenAI were available for this entry. The general idea of recurrent-depth and looped transformers exists in public academic research, but the specific OpenAI implementation, its release status and its access terms are unknown.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Training technique reportedly used for Astra that processes the same text multiple times, obscuring chain-of-thought reasoning. This raises concerns about monitoring and safety.
Closest neighbours by meaning, not just by category.
Model · OpenAI
OpenAI's first reasoning model series, trained with reinforcement learning to think before answering; now deprecated in the API.
Model · OpenAI
OpenAI's GPT-6 Astra, a frontier language model used for reasoning, agentic work and weight interpretation research.
Research · Google
The 2017 Google paper that introduced the Transformer, a neural architecture based solely on attention mechanisms.
Model · DeepSeek
DeepSeek's open-weights reasoning LLM (671B MoE, 37B active) trained with large-scale reinforcement learning.
Model · OpenAI
OpenAI frontier language model used for coding, reasoning and agentic tasks, available via ChatGPT and the API.
Model · OpenAI
OpenAI's open-weight reasoning models, gpt-oss-120b and gpt-oss-20b, released under Apache 2.0.