Anthropic’s international chief says open models carry more risk by definition

Anthropic’s head of international told Bloomberg TV in Dublin that open alternatives to its models carry more risk by definition. That is a broader claim than the one Dario Amodei set out in July, when he wrote that open-weight models without dangerous capabilities are a public good.

Anthropic’s top sales executive outside the US says the open models competing with its own carry more risk. He said it in Dublin, at an Irish government summit.

“We need to be really careful with open-source on safety,” Guillaume Princen, Anthropic’s head of international, told Bloomberg TV on Wednesday. “I can’t talk for open models. But by definition, there’s more risk with them.”

The venue is not neutral ground. The International AI Summit is an official event of Ireland’s EU Council Presidency, opened by Henna Virkkunen. The Irish Council for Civil Liberties has asked Ireland to recuse itself, over its tax reliance on the firms whose files it steers.

Anthropic has a position to protect there. EMEA is its fastest-growing region, Dublin its first EU office, and a Nasdaq listing filed at $965B is close.

The competition is European and recent. Mistral launched Large 4, nicknamed Le Chonk, eight days earlier, a trillion-parameter model trained in Europe and pitched on sovereignty.

The open part has not landed. A preview runs through Mistral’s API and the weights are not due until 27 October, so today it is a hosted model.

Princen went further than his chief executive. Dario Amodei wrote in July that Anthropic has never advocated banning open-weight models, and that those without dangerous capabilities are a public good.

The gap is conditional against categorical. Amodei’s claim turns on whether a model is dangerous. Princen’s turns on the category.

There is evidence for the worry. SaferAI found China’s GLM-5.2 refused none of the offensive-cyber or dual-use-biology tasks. Claude refused so consistently the benchmark could not finish.

Anthropic’s own research this week pointed the same way. Its Frontier Red Team got small models replicating across cheap Android TV boxes, and every model tested was open, including Gemma4 and GLM-5.2.

Openness cut both ways there, since those tests were only possible because the weights were public. Anthropic’s answer is gated access, giving vetted defenders what the model finds rather than the model.

Brussels leans the other way. Article 53 of the AI Act exempts genuinely free and open-source models from part of the documentation owed to the AI Office, unless systemic risk is found.

Princen made a price claim too. He said Claude Opus 5.5 is 40% cheaper than its predecessor. That is the typical-cost figure, since the list price fell 20% and the rest comes from using fewer tokens.