Anthropic AI Updates: June 9, 2026

1. Anthropic Research Argues Biology Databases Need Agent-Ready Retrieval Tools

Anthropic. Anthropic published research finding that AI agents struggle with biological data retrieval because of poorly designed infrastructure rather than reasoning limitations. Testing models on viral sequence queries, the team reports accuracy ranging from 16.9% to 91.3%, but accuracy rose above 90% across all models, reaching 99.7% for GPT-5.5, once they built a deterministic retrieval tool called “gget virus.” The authors argue that biological databases should be redesigned with agent accessibility in mind and that deterministic retrieval tools are currently crucial to making agent workflows reliable. Source

2. Anthropic Releases Swift Package for Claude in Apple’s Foundation Models Framework

Anthropic. Anthropic released a Swift package that lets Apple developers integrate Claude with Apple’s Foundation Models framework. The setup uses on-device models for simple tasks such as summarization, then hands off complex requests involving multi-step reasoning, code generation, and web search to Claude. The integration streams Claude responses back into SwiftUI views and handles streaming, tool calls, and structured outputs automatically. Source

3. Anthropic Adds Observability Dashboard and Directory Submission for Connector Developers

Anthropic. Anthropic introduced a new observability dashboard for developers building connectors, giving them visibility into adoption, errors, and usage across Claude products and by platform. Developers can now also submit their Model Context Protocol servers directly to the connector directory through Claude’s interface. The changes aim to streamline how creators monitor and distribute tools to users who access the directory. Source

4. Claude Code v2.1.169 Adds Safe Mode, a /cd Command, and a Bundled Skills Toggle

Claude Code. Version 2.1.169, tagged June 8, adds a safe mode that launches Claude Code with customizations disabled via --safe-mode or the CLAUDE_CODE_SAFE_MODE environment variable for troubleshooting. It introduces a /cd command that switches a session to a different working directory without breaking the prompt cache mid-session. The release also adds a disableBundledSkills setting and a CLAUDE_CODE_DISABLE_BUNDLED_SKILLS environment variable to hide bundled skills, workflows, and built-in slash commands from the model. Source