AI news, 18 September: Claude now leads a quarter of Anthropic’s own research

Anthropic says Claude leads 26% of its AI R&D. Anthropic reported that its Claude model now “leads” 26% of the company’s internal research and development work on the next generation of AI systems, up from under 1% earlier in the year, according to a company blog post and reports by Reuters and others. “Leads” means the AI can complete most of a task end-to-end from a high-level prompt while a human still supervises; overall, AI collaborates on more than 90% of the work. About 30,000 agents were active at once in August, with every action screened. The company framed the metrics as a way for the public to track progress toward AI that helps build better AI, a key concern in ongoing safety debates. It stressed Claude is not yet fully autonomous in any measured area.

OpenAI launches Astra for Law. OpenAI released Astra for Law, a configuration of its GPT-6 Astra model tailored for legal research, drafting and workflows, Reuters and the company’s announcement confirmed. It pairs the model with a dedicated index of more than 230 million URLs covering U.S. case law, statutes and regulations (drawing on sources including the Free Law Project’s CourtListener), plus specialized instructions. On an internal benchmark it showed roughly 40% relative improvement in correctness over the base model with web search. Initial access is limited to selected law firms via a Trusted Access program in ChatGPT and Codex, with API availability coming; legal-tech firms such as Harvey and Legora can build on it. Integrations with tools like Thomson Reuters are planned.

King Charles meets AI leaders on safety. Britain’s King Charles III hosted senior figures from Nvidia, Google DeepMind, OpenAI and Anthropic at Dumfries House in Scotland, warning of “existential dangers” if the technology falls into the wrong hands and urging safeguards “before it is all too late,” according to Reuters, the New York Times and palace remarks. He said the pace of AI development is “intriguing and deeply concerning in equal measure” and called for it to remain in service of humanity, community and the natural world. Attendees included Nvidia’s Jensen Huang, DeepMind’s Demis Hassabis, OpenAI CFO Sarah Friar and UK AI minister Kanishka Narayan. No binding agreements were expected.

Unsealed filings strain AI fair-use defense. Newly public court documents in the copyright lawsuit brought by the New York Times and other publishers against OpenAI and Microsoft quote executives describing AI products as “largely substitutive” for journalism and an “existential threat” to publishers, Reuters and other outlets reported. A Microsoft director of applied science referred to training on creative work as potentially “an astonishing theft of unprecedented proportions” and the “largest theft of labor in human history.” The outlets argue the quotes undermine the companies’ fair-use claims; Microsoft said the comments reflect one employee’s view and not the company’s position. The case remains ongoing.

Huawei accelerates AI chip push. At its Connect conference, Huawei unveiled the Atlas 960 SuperPoD computing cluster and said it is advancing its Ascend 960 chips, with the 960DT now targeted for the first quarter of 2027—earlier than previously planned—while noting strong domestic demand that outstrips supply, according to Reuters and company statements. The systems aim to support large-scale training and inference and challenge Nvidia’s dominance amid U.S. export restrictions. Huawei also highlighted interconnect technology for linking thousands of processors into larger clusters.