Moonshot AI

Anthropic says it cut what distillers extract from Claude by 90 to 99%, and names Zhipu and Moonshot

The company says reasoning tokens pulled out by distillers are now 10 to 100 times lower than before, and that it has turned away trillions of tokens of paying traffic.

Reading a model's internals instead of its output cut the cost of watching for cyber misuse by about thirty times, Goodfire says

A lightweight probe reads activations at every turn and escalates only what looks wrong, which drops judge cost from $3,166 to $65 per million turns.

Mistral's new trillion-parameter model does the security work Claude and GPT-6 refuse

ML4 is a 1 trillion-parameter open-weight model trained on 3,800 Nvidia chips in Mistral's own European data centres. The number the company leads with is 82% on a vulnerability test where, it says, Claude Opus 5.5 and GPT-6 Astra score near zero because they decline the task.

Reflection says its new open model matches GLM 5.2 on a quarter of the inference compute

The Nvidia-backed lab raised about $4.7bn to put a Western model back on the open-weight frontier. Its first release sits a point below GLM 5.2 on every test it published — and the pitch is the inference bill, not the score.

A German lab put 967 sensitive prompts to six Chinese open models, and one answered 17% of them evenly

An American model was flagged too. Aleph Alpha traced it to about 3,500 rows of a 9.3-million-row fine-tuning set that carried Communist Party framing.

The US will propose a channel with Beijing for telling each other when an AI system goes wrong

Scott Bessent told Axios he raised an incident communication channel with Vice Premier He Lifeng before Xi Jinping's state visit, and that China's own open models are what changed Beijing's mind.