Nathan Lambert puts the open model gap at 4-6 months

Nathan Lambert puts the gap between open and closed AI models at four to six months, citing SemiAnalysis. The estimate comes from his open-models reading list, posted September 11.

Lambert frames open models as a complement to closed ones rather than a replacement. Enterprises deciding whether to build agentic workflows on open weights or closed APIs need a real number, not a vibe. Lambert argues open models won't catch up to closed ones outright. But the gap stays close enough that companies keep building custom workflows on top of them anyway.

That popular explanation for how Chinese labs stay competitive holds that they win mainly by distilling stronger closed models. Lambert says the evidence for that doesn't exist.

He offers no alternative explanation for the pace, only the observation that the distillation story gets asserted, not shown.

The bigger warning is about policy, not capability. Lambert warns vague federal oversight mechanisms could produce a ban on frontier open models within months.