Daily News · 3 min read

Anthropic AI Updates: August 28, 2026

anthropic claude claude-code science robotics legal-tech agents

1. Anthropic Previews the Model Hardware Standard for Lab and Factory Robots

Anthropic. Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared specification that lets AI agents safely discover and operate physical lab and manufacturing devices through a unified interface. The standard introduces standardized drivers so agents can coordinate instruments such as microscopes, liquid handlers, and robotic arms that typically cannot communicate with each other, cutting integration work from weeks to hours. Early partners including Carnegie Mellon, QuEra Computing, Genentech, and the University of Washington reported gains such as 3x faster dose-response experiments and laser recovery success rising from 58 percent to 99.3 percent, with a full open-source release planned after safety evaluations. Source

2. Anthropic Expands Free and Discounted Claude Access for Scientists

Anthropic. Anthropic expanded its support for scientific researchers by opening 10,000 seats offering free and discounted Claude Team subscriptions to scientists worldwide for one year, with premium seats carrying 5x usage limits priced at $15 per month. The company also broadened its AI for Science program beyond biology to additional fields, including compute-heavy research, and offers eligible principal investigators up to $50,000 in credits per project. The move aims to lower cost barriers to advanced AI across academic and nonprofit institutions. Source

3. Claude Code v2.1.248 Adds a Restricted Mode and Usage Credit Requests

Anthropic. Anthropic released Claude Code v2.1.248, which adds a --restricted flag that removes built-in command and code execution tools plus WebFetch while keeping file operations scoped to the working directory. Other additions include an experimental.cacheTtl option to set per-agent prompt cache duration, a /usage-credits command for Enterprise users to request higher limits from administrators, and cross-session messaging via SendMessage and ListAgents between sessions on the same machine. The update also adds server-managed settings diagnostics with a /doctor explanation for load failures. Source

4. Spellbook Runs 530,000 Monthly Contract Reviews on Claude

Anthropic. Anthropic published a customer story on Spellbook, a legal technology platform that automates contract review, drafting, and negotiation for law firms and in-house teams. Spellbook runs roughly 15 Claude agent configurations spanning Opus 5 to Sonnet 4.6 and uses Claude Fable as an oracle model to grade other models, processing about 530,000 contract reviews and 700,000 lawyer chat messages per month. The company reports a 10x efficiency gain that cut review time from about 10 hours to one hour per agreement, serving 5,000 customers across 80 countries. Source

Anthropic. Anthropic published a customer story on EvenUp, a legal technology startup that helps personal injury firms draft demand letters, medical summaries, and discovery responses. The platform uses Opus 4.8 for complex reasoning and Sonnet and Haiku for extraction, reducing document drafting time from 8 to 15 hours down to roughly 30 minutes. Customer firms report 3x higher demand output and a 300 percent increase in settlement offers, with one California firm clearing a 45-day backlog after adoption. Source

6. Rocket Money Builds a Self-Healing Finance Agent on Claude

Anthropic. Anthropic published a customer story on Rocket Money, a personal finance app that built an AI agent called Rowan on Claude to let users manage money through text-based chat instead of dashboards. The team uses a classifier-router architecture with Haiku 4.5 directing requests to specialized subagents running Sonnet 5, with Opus 5 for complex financial reasoning, and designed the agents to self-heal when websites change. Rocket Money reported an 11x increase in monthly code commits over seven months and near-zero hallucinations during beta testing. Source