Claude Sonnet 5 brings Opus-level performance at lower costs, with significant improvements in reasoning, tool use, and coding. The model can plan, use browsers and terminals, and run autonomously at a level that weeks ago required larger and more expensive models.

Claude Science integrates the most common research tools into a single environment, eliminating the need to switch between multiple platforms like PubMed, Jupyter, and terminals. This launch expands Anthropic's focus on accelerating scientific discovery.

Anthropic released details on Fable 5's security classifiers and proposed an industry-wide jailbreak severity framework with Amazon, Microsoft, and Google. This aims to establish unified criteria for evaluating AI vulnerabilities across the industry.

After the U.S. government applied export controls on June 12, Anthropic restricted access to Fable 5 and Mythos 5. Controls were lifted on June 30, allowing global Fable 5 access starting July 1.

Strix is an open-source tool that uses autonomous AI agents for penetration testing, finding vulnerabilities and validating them with real proof-of-concepts. It integrates with GitHub Actions for automated CI/CD scanning.

Simon Willison reports that sqlite-utils 4.0rc2 was developed mostly by Claude Fable for approximately $149.25. This demonstrates the potential of current AI models to contribute to real open-source projects.

A 354 production-ready skills library for Claude Code and 12 other AI coding tools was released, covering engineering, DevOps, security, and compliance. Includes pre-tool security hooks and executive-level personas.

A GitHub issue reports that reasoning-token clustering in GPT-5.5 Codex may be causing performance degradation. The discussion has generated community interest in the internal mechanisms of these systems.

A researcher presented a 10MB LoRA adapter for Qwen3.5-4B that reads internal confidence signals instead of verbalized ones, improving accuracy in tool-use decisions. The system reduces false confidences and private data leakage.

According to reports, Alibaba has classified Claude Code as high-risk software and banned its use among employees. This measure reflects concerns about data security and privacy in AI tools.

As part of a legal dispute with three Hollywood studios, Midjourney is seeking to compel them to reveal their own AI usage details. This reflects a legal strategy around transparency in generative technology use.

The fanfiction community is conflicted over identifying AI-generated content, but detection methods are questionable. Many legitimate authors could be falsely affected by these initiatives.

Herdr is a tool that runs multiple coding agents in a single terminal with visual status of each. Works locally or over SSH without requiring Electron, GUI, or telemetry.

Page Agent enables controlling web interfaces with natural language commands without extensions or headless browsers. Integrates as in-page JavaScript with text-based DOM manipulation.

Taste Skill provides portable skills that improve AI-generated interfaces with better layout, typography, and spacing. Works with ChatGPT, Codex, Cursor, and Claude Code.