Anthropic rejects open-model bans, urges global AI testing

Anthropic’s move to distance itself from blanket bans on open-weight models while doubling down on export controls and mandatory testing is a clear attempt to stake out a middle lane in the AI governance fight. Rather than framing openness itself as the threat, Amodei redirects attention to capability and provenance: who trains the most powerful systems, on whose chips, and under what testing regime. That line is aimed squarely at Washington, which is wrestling with how to respond to rapidly improving Chinese models without accidentally freezing innovation at home.

For the race to AGI, this is a sign that leading labs no longer see pure openness versus closedness as the main axis of competition. The real battle is shifting toward control over compute, distillation pipelines, and who sets the bar for “sufficiently capable” models that must pass pre-release red-teaming. If Anthropic’s position wins out, we are likely headed for a world where both open and closed models coexist, but any system near the frontier must clear standardized safety tests. That could normalize continuous large-scale evaluations as part of the AGI development stack, which in turn shapes how fast new generations can be shipped.

It also sharpens geopolitical lines. By explicitly tying chip export policy to fears of authoritarian AI advantage, Anthropic is aligning itself with a hawkish US posture on China while trying to preserve room for domestic and allied open-weight ecosystems.