Would “pacing the frontier” help or hurt open AI models?

In the September 2026 debate, one critic claimed many enterprises were getting near-frontier performance from reinforcement-learning fine-tuning of open models “20x more cost-effectively,” and alleged closed-model labs wanted to kill that approach with red tape.

Supporters argued that safety costs would fall hardest on the labs with the most capable models. An open-model advocate also argued that shared safety protocols could help adoption rather than threaten it.

The economics drew disagreement, too. One commenter predicted that labs choosing to pace would likely spend slightly more on compute for alignment, monitoring and evaluations, at the cost of lower margins. A reply asked whether keeping costly-to-build models in service longer could instead improve per-model margins.