Anthropic AI Updates: June 19, 2026

1. Anthropic Research Finds Claude Operates Robot Quadrupeds Far Faster Than Humans in Project Fetch Phase Two

Anthropic. Anthropic published Project Fetch: Phase Two, a Frontier Red Team study that revisits a 2025 experiment on whether Claude can help non-experts operate robotic quadrupeds, this time measuring the newest model’s autonomous performance against the original human teams. Operating without human assistance, Claude Opus 4.7 was roughly 20 times faster than the fastest human team across all tasks the participants had completed less than a year earlier, averaging about 37.7 times faster than the unassisted team while producing around 90 percent less code. The model still struggled with precise physical manipulation such as retrieving a beach ball through closed-loop feedback control, which the researchers frame as a remaining capability gap on the path toward physical agentic AI. Source

2. Claude Code Adds Shareable Artifacts for Visualizing Session Work

Anthropic. Anthropic added artifacts to Claude Code, letting users turn a session into an interactive, shareable web page such as a pull request walkthrough, a filterable dashboard, or a release checklist. The pages draw on full session context including the codebase, connectors, and conversation history, update automatically as Claude Code modifies them, and keep version history and galleries for management. Artifacts stay private by default with admin controls for access and retention, and are available in beta for Claude Team and Enterprise organizations through the CLI and desktop app, with pages viewable in any browser. Source

3. Anthropic Launches Enterprise-Managed Authorization for MCP Connectors

Anthropic. Anthropic introduced enterprise-managed authorization for Model Context Protocol connectors, letting organizations control connector access centrally through an identity provider rather than relying on individual user authorization. Users automatically receive connector access on first login based on their IdP group membership, access works consistently across Claude chat, Claude Code, and Cowork, and shorter token lifetimes speed up deprovisioning. The feature is in beta for Claude Team and Enterprise customers, launching with Okta and connectors from Asana, Atlassian, Canva, Figma, Granola, Linear, and Supabase, with early adopters including HubSpot, Ramp, and Webflow. Source

4. Anthropic Publishes a Guide to Steering Claude Code Behavior

Anthropic. Anthropic published a guide explaining seven mechanisms for customizing Claude Code, covering CLAUDE.md files, rules, skills, subagents, hooks, output styles, and an append-system-prompt CLI flag, along with a decision framework for where to place each kind of instruction. The post recommends keeping a root CLAUDE.md under 200 lines and moving procedural content into on-demand skills to conserve tokens and improve adherence. It also cautions that negative instructions in CLAUDE.md are unreliable under pressure and that real guardrails require hooks or managed settings instead. Source