Microsoft Apparently Confirmed OpenAI GPT-6 Uses Looped Transformer Architecture

A Quickly Removed Microsoft Page Sparks Debate

A discussion on Reddit has caught the attention of the AI community: a user reportedly discovered a publicly accessible Microsoft webpage that appeared to confirm OpenAI is using a Looped Transformer architecture in the GPT-6 model series. If accurate, this would lend credibility to earlier reporting from The Information on the same topic.

It's worth emphasizing that the page was subsequently removed by Microsoft — the original poster explicitly noted in an update that "Microsoft updated the web page to remove it." As a result, this information remains at the level of community observation and speculation, without any official statement to back it up. Readers should approach it with appropriate skepticism.

What Are Looped Transformers?

Looped Transformers are an architectural direction that has attracted growing interest in model research. Unlike standard Transformers, where data flows through each layer in a single forward pass, a looped architecture allows the model to iterate multiple times over the same weight layers — effectively increasing the model's "computational depth" without significantly increasing its parameter count.

The core idea is to trade compute for reasoning capability — by reusing the same set of weights across multiple processing rounds, the model can achieve performance on complex reasoning tasks comparable to a much deeper network, while keeping the number of stored parameters under control. For next-generation large models targeting stronger reasoning ability, this is a technically compelling path.

The academic roots of Looped Transformers trace back to theoretical research on the expressive power of Transformers. Researchers found that the number of layers in a standard Transformer caps the number of computational steps it can simulate. A looped, weight-sharing structure theoretically allows a model to simulate arbitrarily deep computation graphs — approaching Turing completeness. Several papers from institutions like DeepMind and CMU (e.g., Universal Transformers) have validated this direction: on tasks requiring multi-step reasoning such as algorithmic problems and math questions, looped architectures consistently outperform standard Transformers with the same parameter count.

Notably, the number of loops can be dynamically adjusted at inference time, meaning the same model can "allocate compute on demand" based on task difficulty — breezing through simple problems and iterating more on complex ones. This property closely aligns with the emerging paradigm of Adaptive Computation, and is precisely why the architecture is seen as a solid foundation for building model variants across different capability tiers.

Key Technical Details from the Post

Based on the Reddit post, several noteworthy technical points stand out:

• GPT-6.1 Sol uses 2 inference passes. The post also mentions the phrase "instead of three," implying different model versions may use a different number of loops.
• Regarding the phrase "same base model weights as GPT-6 Sol," the poster offered their own interpretation: Microsoft likely means that both GPT-6 and GPT-6.1 are post-trained on top of the same pre-trained base model, rather than sharing identical final weights.
• In other words, GPT-6 and GPT-6.1 Sol share the same pre-training foundation but use different post-training pipelines and differ in loop count (GPT-6.1 uses one fewer loop).

If this interpretation holds, the number of loops may become a tunable dimension for OpenAI to ship different product variants from a single base model — flexibly balancing compute cost against capability by simply changing the number of inference-time iterations.

Why This Rumor Is Worth Paying Attention To

From an industry perspective, the significance of this rumor lies in what it may reveal about how leading labs are actually approaching architectural evolution. For a long time, improvements in large model capability were driven primarily by scaling up parameters and training data. Looped architectures represent a different approach: investing more compute at inference time in exchange for stronger reasoning performance.

This echoes the broader trend of "test-time compute" that has gained momentum in recent years. If OpenAI has indeed deployed Looped Transformers in the GPT-6 series, then the strategy of using a single base model with varying loop counts means a provider can flexibly configure models for different scenarios — a lightweight version optimized for speed, a flagship version pushing maximum capability — without needing to train multiple independent models from scratch.

"Test-time compute" refers to the strategy of investing more computational resources during inference rather than training, in order to improve output quality. OpenAI's o1/o3 series has already demonstrated the effectiveness of this approach by extending Chain-of-Thought generation. Looped Transformers can be seen as a native architectural implementation of the same idea — the difference being that Chain-of-Thought scaling adds compute along the token sequence dimension, while looped architectures repeatedly process the same hidden state along the network depth dimension. The two are not mutually exclusive and could theoretically be combined. If GPT-6 does use a looped architecture, it would suggest that OpenAI has baked "scalable test-time compute" directly into the model design itself, rather than relying solely on prompt engineering or decoding-level tricks.

Stay Grounded: This Remains Unverified Community Observation

It bears repeating: all information in this article originates from a single Reddit post, its core piece of evidence (the Microsoft webpage) has been deleted and cannot be independently verified, and the poster's interpretation of phrases like "base model weights" is personal speculation.

Until OpenAI or Microsoft provides an official statement, claims about GPT-6 using a Looped Transformer architecture should be treated as an unconfirmed industry rumor, not established fact. That said, community discoveries like this often serve as early signals of emerging technical trends. The underlying architectural directions they point to — looped computation and test-time scaling — are genuine focal points in current large model research, and well worth tracking going forward.