Research published in early August found that open-weights models have been catching up to frontier models in capability, while the distance in safeguards remains.
Converging capability is the data point that changes the market. If the quality gap narrows, the case for selling access by subscription weakens, which explains the period's moves, with companies publishing weights of their latest models.
The safeguard gap is the data point that changes the risk. A published model can be tuned by anyone, including to remove the restrictions it shipped with, and once published it doesn't come back.
Both together describe the impasse the sector discussed all month, across a coalition letter, a published position from a company that refused to sign, and a voluntary standard negotiated with government.
For anyone choosing a model, the practical consequence is that evaluation can't stop at capability benchmarks. It's worth asking what the model does when handed a request it should refuse, and testing that on your own workload.