Anthropic previews Model Hardware Standard letting Claude control robots and lab tools

Anthropic unveiled a Model Hardware Standard (MHS) on August 27, a software layer that lets Claude interface with robots and scientific lab tools, alongside an expansion of its support for scientists. It marks Anthropic's first structured move into embodied and lab-automation AI, an area historically dominated by robotics-specific stacks rather than general LLM providers.

Crucially, Anthropic is shipping MHS as a research preview rather than a full open standard. On X, the company said 'there's more to learn before we open source MHS,' noting that LLMs 'still lack physical intuition, having learned about the physical world from text and images,' and that the preview period will let it build more safety evaluations and strengthen protections for using AI in physical settings. That caution is consistent with Anthropic's broader safety posture — it recently raised its internal catastrophic-misalignment risk from 'very low' to 'low' and has withheld an unreleased 'Model 2' pending safety work.

Mechanically, a hardware standard implies a defined interface for perception, planning and actuation — letting Claude issue tool calls that map onto robot controllers or lab instruments (liquid handlers, microscopes, synthesis rigs). For scientific workflows, that could turn Claude into an orchestration layer over automated wet-labs, a direction Google (Isomorphic) and others have pursued from the biology side.

Competitively, this pits Anthropic against NVIDIA's physical-AI push (Cosmos 3 world models, robotics on SageMaker) and Google DeepMind's robotics work, but from a language-model-first angle. The skeptical read, voiced by researchers like Ethan Mollick around the concurrent METR agent report, is that ascribing capability or 'intuition' to LLMs in the physical world invites overreach; Anthropic's own 'lacks physical intuition' framing pre-empts that. Watch which lab partners get preview access and whether Anthropic publishes physical-world safety evals before opening the standard.