What's new in Claude

Added /config key=value syntax for setting any config from the prompt, improved long-paragraph streaming, enhanced auto-retry for mid-thinking API drops, and fixed prompt caching on custom API URLs, file truncation on network drives, and macOS TUI freezing.

Fixed mid-stream connection drops and restored mouse-wheel scrolling in WSL2.

Added Tool(param:value) syntax for permission rules and improved nested .claude/ directory handling.

/cd command: move the active session to a new working directory mid-conversation without rebuilding the prompt cache — no more restarting to switch projects.

Sub-agents can spawn sub-agents: background chains now support up to five levels of nesting, enabling fully automated multi-tier orchestration.

--safe-mode: launch Claude Code with all customizations, plugins, and hooks disabled for clean troubleshooting.

fallbackModel: configure up to three fallback models tried in order when the primary model is unavailable or rate-limited.

Claude Sonnet 4 (claude-sonnet-4-20250514) and Claude Opus 4 (claude-opus-4-20250514) are now retired — all requests return an error. Upgrade to Sonnet 4.6 or Opus 4.8.

The GET /v1/environments/{id}/work endpoint for self-hosted sandboxes is now available on Claude Platform on AWS, authorized via the new GetEnvironment IAM action.

Claude Fable 5 launched as the most capable widely available model — 1M token context, 128K output, always-on adaptive thinking, and new stop_reason: "refusal" with zero billing on pre-output refusals.

Claude Managed Agents now supports scheduled deployments (cron-style) and environment variable credentials in vaults for secure secret injection into agent sandboxes.

Doubled rate limits on Claude Code and raised API limits for Claude Opus so developers can build more reliably at scale — both changes now live. Programmatic usage credits (covering Agent SDK, claude -p, GitHub Actions) kicked in June 15.

Anthropic opened a Seoul office with new partnerships across the Korean AI ecosystem — the latest expansion as Claude goes global.

One year since Claude Code GA: Boris Cherny hasn't written a line of code by hand in 8 months. "Some days I'm coordinating tens of thousands of agents." Claude-to-Claude prompting now dominates — humans are increasingly the exception. Code output inside Anthropic is up 8x since January 2026. (Fortune, June 11)

Dynamic workflows shipped: Claude now writes its own task harness on the fly, custom-built for each job, orchestrating dozens to hundreds of subagents from a single script. Also: Todos are becoming Tasks in Claude Code for richer state tracking.