Claude Code v2.1.208 Major Updates - Access to Claude Fable 5 Re-extended Until July 19

This is Ishikawa from the Cloud Business Division. Claude Code v2.1.208 (released July 13, 2026) has been released. This release focuses on stability improvements, including the addition of new features such as screen reader support, along with memory optimization for long sessions and numerous bug fixes.

Update Summary

v2.1.208 includes 45 changes. There are 4 new features, with the remainder focused on performance improvements such as resolving memory leaks and significantly reducing transcript sizes, along with more than 30 bug fixes. There are no breaking changes or deprecations. The release also includes behavioral improvements around background agents and a fix for a Bedrock authentication regression in AWS SSO environments.

On a recent note (7/13), access to Claude Fable 5 has been re-extended until July 19 for all paid plans.

https://x.com/claudeai/status/2076351399999557669

Notable Updates

• New Feature: Added a plain text display mode for screen readers. (Enable with --ax-screen-reader / CLAUDE_AX_SCREEN_READER=1 / setting axScreenReader.)
• Bug Fix: Fixed an issue where resuming a long-context session after an automatic update would incorrectly display "context usage 100%." Additionally, a regression in 2.1.207 where Bedrock authentication failed for AWS SSO profiles where sso_region differs from the Bedrock region has also been fixed.
• Security: Destructive deletions (e.g., rm -rf ~) containing command substitutions such as $(…) or backticks will now prompt for confirmation even in --dangerously-skip-permissions or auto mode.

Update Details

New Features

• Screen Reader Mode: A "screen reader mode" has been added, providing opt-in plain text rendering for screen reader users. There are three ways to enable it:

  1. CLI flag: claude --ax-screen-reader
  2. Environment variable: CLAUDE_AX_SCREEN_READER=1
  3. settings.json: "axScreenReader": true

  The ax stands for accessibility, which is reflected in the flag name. Writing it in settings.json means you don't need to add the flag every time, making the settings file approach convenient for users who always want it enabled.

• vimInsertModeRemaps Setting: In vim mode, you can now map 2-key insert mode sequences like jj to keys such as Escape.

• CLAUDE_CODE_PROCESS_WRAPPER Environment Variable: Agent views and background services can now route all of Claude Code's own process launches through a wrapper executable specified by the enterprise. This is a feature designed to accommodate enterprise launcher requirements.

• Fullscreen Mode Mouse Operations: Mouse click support has been added for multi-select menus and "Other" input rows.

Performance & Memory Optimization

This release is notable for memory and disk usage optimizations in long-running and large-scale sessions.

• Transcript Size Reduction: Session transcript sizes have been reduced by up to 79x in heavily edited sessions, and checkpoint disk usage is now limited by cleaning up overwritten file history backups.
• Faster MCP Tool Invocation: For print/SDK sessions with many MCP tools, tool pool assembly is now cached, reducing CPU overhead per tool call. With a large number of tools, tool rounds can be up to 7x faster.
• Improved Input Responsiveness: Improved input responsiveness by not redrawing the entire UI during agent task list updates.
• Edit Cache Memory Cap: Changed the file edit read cache from pinning up to 1,000 files to a 16 MB cap, reducing memory usage.
• Reduced Memory on Session Resume: Reduced memory usage when resuming sessions that include background agents or forks generated from large conversations.

Improvements

• Display of Completed Background Agents: Completed background agents now remain visible in /tasks until cleanup, rather than disappearing the moment they finish.
• Connecting to Stopped Agents: When connecting to a stopped background agent, the transcript is now shown immediately while the session warms up, rather than displaying a blank "Session is starting" screen.
• Ctrl+X in Agent View: Ctrl+X now deletes the worktree for renamed branches, does not discard unpushed commits, retains the session row if the worktree is kept, and resets reused worktree names to the current base.
• /usage Fallback Display: When the usage endpoint is rate-limited, /usage now shows the most recent usage bar with an "as of" note instead of an error screen.

Security

• Confirmation for Destructive Deletions Containing Command Substitution: Destructive deletions (e.g., rm -rf ~) in commands containing $(…) / backticks / <(…) will now prompt for confirmation in --dangerously-skip-permissions and auto mode, just as plain forms do. This closes a loophole where command substitution could previously bypass the check.

Fixes

• Fixed False "100%" Display After Auto-Update: Fixed a bug where the context window (and auto-compact indicator) was temporarily reset to 200k after a CLI auto-update, causing long-context sessions to incorrectly display "context usage 100%" upon resumption.
• Fixed Bedrock (AWS SSO) Authentication Failure: Fixed a bug where Bedrock authentication would fail with "Session token not found or invalid" for AWS SSO profiles where sso_region differs from the Bedrock region (a regression in 2.1.207).
• Fixed Edit Tool Failures: Fixed a bug where the Edit tool would fail on files that had been modified after being read, even when the target text matched uniquely.
• Fixed Various Read / Grep / Glob Issues: Fixed Read incorrectly reporting empty files as "shorter than offset," Grep silently returning "No files found" for invalid regular expressions, Grep's count mode underreporting totals during pagination, and Glob crashing with an unclear error when paths or similar contained NULL bytes.
• Fixed Crash on HTTP/2 Disconnection: Fixed a bug where supervised/background sessions would crash when a server terminated an HTTP/2 connection with GOAWAY during request processing.
• Fixed Lost Replies to Background Agents: Fixed a bug where replies entered to a background agent would be lost on delivery failure. Text is now saved and delivered when the session restarts.
• Fixed Missing Output in claude -p: Fixed a bug where stream-json/JSON output would be truncated and result messages would be missing when piping large responses with claude -p.
• In addition, numerous minor bugs have been fixed, including memory leaks, multi-second delays caused by permission rule evaluation, auto-restoration of fast mode, context contamination in /release-notes, display issues in /usage and /mcp, and stream-json input stability.

Closing

v2.1.208 is a release that includes a new screen reader support feature while focusing primarily on memory optimization for long sessions and overall stability improvements. In particular, the false "100%" display after auto-updates and the Bedrock authentication failure in AWS SSO environments (a regression in 2.1.207) are fixes that are likely to affect users on a daily basis. We recommend that users running Claude Code via Bedrock update as soon as possible. Why not update and give it a try?

References

https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md

https://code.claude.com/docs/en/changelog