Anthropic names first Head of Claude for Legal
The legal industry is becoming AI's next battleground: Anthropic hired a legal AI founder to run Claude's push into law, a Chinese robotics startup claims a sorting record that beats Figure AI at a fraction of the hardware cost, and AT&T says open models will soon power the majority of its AI.
Anthropic has hired Robert Mahari, founder of the legal AI startup Akiva AI, as its first "Head of Claude for Legal" — a sign that the company intends to turn legal work into a serious commercial push.
Mahari, a Stanford CodeX fellow with a doctorate in legal AI from MIT, announced the move on LinkedIn this week. He joins product lead Mark Pike on the team behind Claude for Legal, the vertical Anthropic launched in May with twelve plugins for legal work and partnerships with more than twenty legal-tech companies, including LexisNexis and Relativity. In his announcement, Mahari said he will focus on showing law firms and in-house teams what the product can do, and noted that Anthropic's models top Legora's recent benchmark for agentic legal reasoning.
The hire is part of a broader land grab: OpenAI brought on Ironclad founder Jason Boehmig, Microsoft shipped a legal agent inside Word, and Amazon launched Quick for legal tasks. The telling pattern is that the labs aren't just licensing models to law firms anymore — they're hiring practitioner-founders to build trust and go-to-market muscle in a profession where confidentiality decides adoption.
A Chinese embodied-AI startup called 自变量机器人 (X2 Robot) says its robots just set a global logistics-sorting record, beating Figure AI's published benchmark by 45% at roughly a third of the hardware cost.
The company livestreamed a one-hour sorting test in which a dual-arm robot with standard grippers processed 1,816 packages an hour — random-shaped parcels on a moving belt, no scripted motions, no human takeover — versus the 1,248-per-hour average Figure AI has published for its humanoid setup. Accuracy hit 98%. The company credits its WALL-B "world unified model," which fuses vision, language, touch and action in a single network rather than routing data between separate modules. The hardware contrast is the point: dual arms and industrial grippers cost roughly 70% less than a humanoid with dexterous hands.
It's a company-claimed live demo, not a third-party eval, and the comparison isn't perfectly apples-to-apples — different hardware, different parcel mix. But the direction matters: Chinese labs are increasingly betting that a smarter "brain" can substitute for expensive hardware, which is exactly the cost curve that makes embodied AI deployable at warehouse scale.
AT&T says open-weight models already power about a quarter of its AI usage — and it expects them to handle 70% to 80% over time.
The telecom runs an average of 45 billion AI tokens a day, according to a Wall Street Journal profile of Chief Data and AI Officer Andy Markus, who says open models have cut costs by 80% to 90% in certain applications. AT&T built a "smart router" that picks the cheapest adequate model per task, runs open models on its own infrastructure, and cites data control — "AI sovereignty" — as the strategic reason to avoid depending on any single proprietary vendor. Gartner projects open models will underpin more than half of business AI use cases within two years, up from under 10% today.
The open-versus-closed debate usually gets fought on benchmarks, but the enterprise economics are quietly decisive. When a company burning 45 billion tokens a day says open weights are its future, that's a data point the frontier labs can't ignore.
What to watch: whether OpenAI's rumored legal offering lands before Anthropic's legal team gets much further — and whether any frontier lab answers AT&T's open-model math with cheaper proprietary tiers.
Do you think most enterprises will follow AT&T toward open-weight models, or does proprietary quality still win the day? Tell us in the comments.
