What this is

This week, Reddit users noticed that a Microsoft public page briefly acknowledged that OpenAI's GPT-6 series uses a "Looped Transformer" architecture — in plain terms, an approach that has the model run two or three inference passes on the same query before producing an answer, rather than stopping after a single pass. According to the screenshots, GPT-6.1 Sol runs 2 loops while GPT-6 runs 3 loops, but they share the same underlying pretrained weights — the differences are in post-training and loop count.

A few hours later, Microsoft edited the page and the information was removed. But quick-eyed users had already captured screenshots, and The Information's earlier reporting was retrospectively corroborated as a result.

Industry read

What's most worth unpacking here isn't the Looped Transformer itself — the academic community has been discussing this line of thinking for two years, and Meta and others have published related papers. What's notable is OpenAI's unusual posture of disclosure.

One read: OpenAI is being pushed toward transparency by competitive pressure. After GPT-5's lukewarm market reception, with Anthropic and Google continuously taking share in vertical use cases like coding and agents, OpenAI needs stronger technical narratives to justify each generation's price tag. The looped architecture is a strong story — it means GPT-6 isn't just stacking parameters, but trading inference count for quality. A smarter narrative.

But a cooler reading: this may simply be leftover technical documentation, or an unreviewed marketing draft. Microsoft pulled it fast — which tells us both sides realized this wasn't a prepared official statement. Until OpenAI explains it themselves, treating this as a "confirmed technical direction" would be premature.

Impact on regular people

  • For enterprise IT: If GPT-6.1 truly "runs twice before answering," the single-inference cost structure will shift. Procurement teams may need to recalculate billing logic and SLA evaluations.
  • For individual professionals: If AI takes one or two seconds longer to answer when you're drafting emails or summaries, don't rush to complain — the model may simply be "thinking one more step," and output quality will be higher.
  • For the consumer market: No visible changes in the short term. Whether OpenAI turns "loop count" into a marketing hook (e.g., "think three times before acting") is worth watching.