Anthropic flags GLM-5.3 as open-model debate heats up

The report lands in the middle of a fight that has been public for some time. Nvidia CEO Jensen Huang, who gains if people keep buying hardware to run open models, pushes for the United States to develop more of them, while Anthropic's Dario Amodei, whose closed-model business could be undermined by cheap and powerful alternatives, takes a more cautious stance. Those commercial incentives sit on top of a real disagreement that experts describe as downstream of how you view AI itself.

Today closed models lead — as of publication the top eight models on the Artificial Analysis LLM leaderboard are all closed — but open models are only a handful of months behind. The defensive tools that apply to a closed model, filters and account shutdowns, do not travel with a downloaded model, which is why the debate has moved away from an all-or-nothing framing toward techniques meant to reduce the risk. Those include training models on data stripped of bad behavior and building models resistant to retraining, though a recent preprint found no current tamper-resistant approach survives every attack.

The benefits are not as clean as they first appear either. Open weights are described as a bare minimum rather than full openness, and without affordable chips the weights cannot be run at all. Several experts point to staged release, where a model starts closed and opens up as its safety is assured, as a way to hold the trade-off between safety and deconcentration of power.