Anthropic will make Auto Mode the default in Claude Code for Pro, Max, and Team plans starting August 14, shifting from manual approval of code edits to automated safety filtering (The Decoder). In internal testing, the Auto Mode classifier caught 89 percent of dangerous commands compared to 13.6 percent caught by human reviewers (The Decoder). Separately, Claude Code sessions can now communicate with each other and share context across parallel terminal instances on macOS and Linux (The Decoder).
Climate scientist Zeke Hausfather tracked his eight-week Claude Code usage at 3.2 billion tokens and approximately 170 kilowatt-hours of data center electricity, finding that agentic workloads consume roughly 600 times more energy per interaction than typical chat prompts (The Decoder). The measurement highlights a gap between token-level energy reporting published by cloud providers and actual infrastructure demand in agent-based workflows. OpenAI also paused parts of development on its Astra model after internal testing showed cybersecurity capabilities strong enough that the company can no longer rule out its highest risk level under its safety framework (TechCrunch).