US Lifts Restrictions on Anthropic AI Models

Highlight

Claude Sonnet 5 Is Now the Default Model

Sonnet 5 is now the default in Claude Code, with a native 1M-token context window and promotional pricing of $2/$10 per Mtok through August 31. You should notice better output quality on delegated tasks too: subagents and context compaction now inherit your session's extended thinking configuration. Sonnet 5 sessions also skip the mid-conversation system role that previously delivered harness reminders. See the Sonnet 5 announcement for details.

Worth Knowing

Subagents Now Run in the Background by Default

Claude keeps working while subagents run and notifies you when they finish, rather than blocking the turn. Background sessions in claude agents also fire Notification hooks (agent_needs_input / agent_completed) when they need input or complete.

Claude in Chrome Reaches General Availability

Claude in Chrome is now GA. Plan mode also now correctly prompts for state-changing browser tool calls while auto-allowing read-only browser_batch calls.

Default Permission Mode Renamed to Manual

The permission mode formerly called "default" is now labeled "Manual" across the CLI, --help, VS Code, and JetBrains. Both --permission-mode manual and "defaultMode": "manual" are accepted, and the previous default value still works.

Background Agents Auto-Open Draft PRs

Background agents launched from claude agents now commit, push, and open a draft PR when they finish code work in a worktree, rather than stopping to ask. Session rows show pull-request links as bare #N in the agents view.

AskUserQuestion Dialogs No Longer Auto-Continue

AskUserQuestion prompts now wait for your input by default instead of timing out and continuing on their own. If you preferred the old behavior, you can opt into an idle timeout through /config.

Admins Can Set an Organization Default Model

Admins can now set a default model in the org console. When you haven't picked one yourself, /model shows it as "Org default" or "Role default."

Under the Hood

Background Agent Stability Fixes

A batch of fixes stops the Linux daemon from killing itself every ~50 seconds after an unclean shutdown, prevents claude stop from being undone by respawns, keeps older reinstalled builds from taking over the daemon, resumes workers killed by daemon restarts, and stops sessions from silently stalling after sleep/wake. macOS sessions no longer show "Reconnecting…" every ~52 seconds in the agents view.

Better Handling of Transient Network and Streaming Errors

Brief network drops mid-response (ECONNRESET) now retry with backoff instead of aborting the turn, transient 429s unrelated to your usage limit are retried automatically for subscribers, and mid-stream server errors keep the partial output with an incomplete-response notice. TLS certificate errors now fail immediately with actionable guidance instead of burning retries. /code-review also cuts token usage by roughly 25% by merging five cleanup finders into one.

From Anthropic

In the News