Report: OpenAI diverts 10% of compute to safety as loop architecture complicates next flagship release

141_1337 · reddit · 2026-10-14

Per The Information, OpenAI is racing to make its next flagship model safe to release after canceling a less capable model over safety concerns.

The successor uses loop transformers that repeatedly process problems internally, making reasoning harder to monitor
Following the Hugging Face incident, OpenAI redirected roughly 10% of its compute toward alignment and safety
The model will undergo US government testing for up to 30 days pre-release, per an executive order on advanced AI evaluations
Employees still expect one or two new models by year-end; GPT-6.1 Sol shares GPT-6 Astra weights but uses two inference passes instead of three
OpenAI is also contending with fallout from firing 3 safety researchers

More from Models