Anthropic AI Updates: June 6, 2026

1. Anthropic Tests Claude on Chemistry and NMR Spectroscopy

Anthropic. Anthropic published research evaluating how Claude performs on chemistry tasks, with a focus on nuclear magnetic resonance (NMR) spectroscopy used for molecular structure determination. The company reports that Opus 4.7 predicts NMR spectra from molecular structures comparably to specialized software such as ChemDraw and MestReNova, and can also propose structures in reverse from experimental data. The post outlines planned improvements in reading and rendering chemical structures, reaction planning, and literature comprehension as part of Anthropic’s AI for Science program. Source

2. Anthropic Publishes Claude Cowork Product Guide

Anthropic. Anthropic released a product guide for Claude Cowork, positioning it as a knowledge work agent that carries multi-step tasks through to finished deliverables rather than requiring manual follow-up. The guide notes that Cowork reads and writes local files and works across connected applications such as Slack and Google Drive. It covers setup, workflows, and how to extend Cowork with plugins for automating knowledge work. Source

3. Anthropic Details How a Non-Technical Seller Built Tools With Claude Code

Anthropic. Anthropic published a case study describing how Jared Sires, a sales representative with no coding background, used Claude Code to build an email-drafting tool called CLAFTS that he reports saved him two to three hours per day. The success prompted a broader effort to develop Claude-powered tools across the sales organization. Sires’s role shifted toward go-to-market product management focused on automating team workflows. Source

4. Claude Code v2.1.166 Adds Fallback Models and Glob-Based Deny Rules

Claude Code. Version 2.1.166, tagged June 6, adds support for configuring up to three fallback models that are attempted sequentially when the primary model is unavailable, including a --fallback-model flag for interactive sessions. The release also adds glob pattern support in deny rules so wildcards can match tool names, expanded thinking controls to disable extended thinking via MAX_THINKING_TOKENS=0 and --thinking disabled, and URL filtering in the agents session list. It additionally hardens cross-session messaging so messages relayed via SendMessage no longer carry user authority. Source