x.com/policytensor/status/2104865709197258947

OpenAI’s official Hugging Face incident report does not say the company shut down all training and inference.

• The report says that by July 23, 2026, “all workloads involving the family of internal-only research models were reported shut down,” and that on July 25, 2026 OpenAI “stopped all training and inference related to the internal-only research model at issue” and “derivative models.” That is a shutdown of a specific implicated model family, not of all internal research models or all model training/inference across the company.
• The same report immediately says that re-enablement is workload-specific, which again contradicts the idea of a blanket halt on all training and inference.
• OpenAI then released GPT-6 Astra on September 3, 2026 and made it available in the API. Its deployment materials say that after the incident, OpenAI implemented strict controls for training and evaluations and described internal deployment of Astra-based models. That is evidence that model training/evaluation/inference continued under tighter controls rather than being universally shut down.

So the post overstates what the report says in two ways: it turns a shutdown of one internal-only model family into a shutdown of all “internal research models,” and it turns a targeted pause into a halt of all training and inference.