OpenAI’s next flagship model uses loop transformers, a design that runs a problem through the model several times and can make its reasoning harder for humans to monitor, according to The Information. OpenAI has never acknowledged using the technique, but Microsoft disclosed last week in a blog post that GPT-6.1 Sol uses GPT-6 Sol’s base weights with two passes instead of three, then deleted the passage. OpenAI didn’t comment on loops, but a memo from its research leaders said keeping frontier models monitorable will take a commitment from the whole industry, OpenAI included. Getting the new model ready could take a while, because once OpenAI is satisfied with its safety results, the U.S. government gets up to 30 days to test it before any partners see it.
