US Bars Foreigners From Top Anthropic Models

Highlight

Claude Fable 5 Lands, Then US Restricts Access

Version 2.1.170 added access to Claude Fable 5, a Mythos-class model Anthropic positioned as its strongest generally available release. Within three days, a US government export directive forced Anthropic to suspend access to Fable 5 and Mythos 5; if you previously routed work to Fable 5, auto mode now falls back to the best available Opus model, and a new classifier fallback cookbook covers detecting safety-classifier blocks and falling back to Opus 4.8.

Worth Knowing

Sub-Agents Can Now Spawn Sub-Agents

Your sub-agents can now spawn their own sub-agents, up to five levels deep, letting you decompose work into deeper trees without orchestrating everything from the top-level session.

Safe Mode Flag for Troubleshooting

You can now launch with --safe-mode (or CLAUDE_CODE_SAFE_MODE) to disable CLAUDE.md, plugins, skills, hooks, and MCP servers in one shot. Useful when you need to isolate whether a problem comes from your config or from Claude Code itself.

New /cd Command Moves Sessions Between Directories

The new /cd command moves an active session into a different working directory without invalidating the prompt cache, so you can pivot to another repo or worktree mid-conversation without paying the cache-miss tax.

Managed availableModels Now Constrains Defaults

If you administer Claude Code policies, the new enforceAvailableModels managed setting makes the availableModels allowlist also bind the Default model, and prevents user or project settings from widening a managed list. A Default that would resolve to a disallowed model now falls back to the first allowed one.

VS Code /usage Adds Per-Skill and Agent Breakdowns

The VS Code /usage dialog now breaks down consumption by cache misses, long context, subagents, and per-skill, agent, plugin, and MCP usage over the last 24 hours or 7 days, making it easier to see where your spend is actually going.

Session Titles Match Your Conversation Language

Session titles are now generated in the language you're actually working in. Set the language setting if you want to pin titles to a specific language regardless of conversation content.

Under the Hood

Lower CPU Use in Long Conversations and While Idle

Long conversations skip redundant message normalization and full history transforms when streaming tool-use state is unchanged, the /goal chip no longer redraws the terminal at 5 Hz while idle, and a 30-50 ms per-turn UI stall for macOS claude.ai users is gone.

Model Picker and Bedrock Region Fixes

The /model picker now correctly shows the family that Default resolves to, stops offering models Bedrock doesn't serve, and respects availableModels for subagent overrides, the advisor model, and version-specific IDs. Bedrock now reads ~/.aws config files when AWS_REGION isn't set, and GovCloud regions no longer derive a bad global inference profile prefix.

From Anthropic

In the News