What's new in Claude

Added sandbox.credentials setting to block sandboxed commands from reading credential files and secret environment variables; also includes fixes for structured output, MCP tool timeouts, and multiple UI/stability improvements.

Added claude mcp login/logout CLI commands for MCP server authentication without interactive menus; MCP OAuth browser styling improvements; fixes for subagent management, permission prompts, and background session handling.

/cd command: Move the current session to a new working directory mid-conversation without rebuilding the prompt cache — useful when switching between repos in a long session. (Week 26 digest not yet published.)

Sub-agents can spawn sub-agents: Background chains now supported up to five levels deep, enabling more complex multi-agent orchestration patterns out of the box.

--safe-mode flag: Starts Claude Code with all customizations disabled (hooks, plugins, settings overrides) for clean troubleshooting when something in your config is misbehaving.

fallbackModel config: Configure up to three fallback models tried in order when your primary model is unavailable or rate-limited, keeping long-running sessions from failing silently.

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

Tool version updates: Code execution tool adds code_execution_20260521 with an explicit 90-second per-cell time limit. Web search and web fetch add response_inclusion to drop consumed result blocks in agentic API responses. No beta header required for either.

Claude Fable 5 launched: Most capable widely-released model — 1M token context window, 128k max output tokens, always-on adaptive thinking, and a new tokenizer (~30% more tokens vs. pre-Opus 4.7 models). Requires 30-day data retention; not available under zero data retention. Claude Mythos 5 also launched for Project Glasswing participants.

Claude Tag launched: Always-on AI teammate that lives in your Slack channels. Claude monitors assigned channels and intervenes proactively with reminders, summaries, or context — without waiting to be tagged. Available in beta for Team and Enterprise. Internally, 65% of Anthropic's product team's code is now written by their internal version of Claude Tag.

Thariq on Claude Tag: "Claude Tag is an incredible new form factor for agents." Also: Claude Code can now upload and edit HTML artifacts shareable with your team or other Claudes — starting with Team plans, coming to Pro and Max soon.

Boris (Claude Code creator) on his setup: "Claude Code works great out of the box — there is no one correct way to use it." His #1 tip: turn on auto mode. "No more permission prompts — the key building block for multi-Clauding." Also noted he hasn't written a line of code himself in 2026.