Claude Code v2.1.236~v2.1.237 Major Updates - Addition of ANTHROPIC_DEFAULT_MODEL and Output Style Concise

This is Ishikawa from the Cloud Business Division. I'd like to introduce a summary of the Claude Code v2.1.236 ~ v2.1.237 updates (both released on August 19, 2026). These two versions include the addition of a new output style and a fix for cache-related issues when using a gateway, making them worth reviewing your settings for.

The previous update article is here.

https://dev.classmethod.jp/articles/20260819-cc-updates-v2-1-235/

Update Summary

This covers 2 versions (v2.1.236 ~ v2.1.237, released August 19, 2026), with a total of 35 changes. The breakdown is 4 new features, 3 security updates, 11 improvements, and 17 fixes, with nearly half being bug fixes. Multiple changes also relate to permissions around sandboxes and auto mode.

Notable Updates

ANTHROPIC_DEFAULT_MODEL to specify the starting model for new sessions (v2.1.236)

The environment variable ANTHROPIC_DEFAULT_MODEL has been added. It specifies the model that new sessions start with, and if a selection is made via /model, that takes priority and the selection is retained across restarts. This behavior differs from ANTHROPIC_MODEL.

For those who want to align the starting point of new sessions across a team, this feels like a more manageable option than locking things down with ANTHROPIC_MODEL.

notify_when_idle to notify once when another session becomes idle (v2.1.236)

notify_when_idle has been added to cross-session SendMessage. It allows you to request that another Claude Code session on the same machine send a single notification the next time it becomes idle. It is opt-in and one-shot, with no polling (macOS / Linux).

For those running multiple parallel sessions with long-running processes, this feels like a change that makes it easier to use waiting time for other work.

Fix for multiple features breaking after a directory is deleted (v2.1.236)

A bug has been fixed where clipboard copying, background housekeeping, background sessions, and local MCP logging would stop working after the directory a session had moved to was deleted. This was occurring since v2.1.229.

If you move directories during a session and have been experiencing issues since v2.1.229, it's worth checking whether this version resolves the problem.

macOS sandbox read-deny rules now take priority even inside allowed areas (v2.1.236)

For the macOS sandbox, wildcard read-deny rules (e.g., **/.env) now take priority even inside read-allowed areas. Directory contents that match are also covered, and it is no longer possible to bypass this by renaming the targeted files.

For those using the macOS sandbox to protect files containing secrets, it's worth confirming that intended denials are now working more reliably.

Monitor allow rules temporarily disabled while in auto mode (v2.1.236)

While auto mode is enabled, Monitor allow rules are now temporarily disabled. This means Monitor commands are also confirmed in the same way as Bash commands.

For those who regularly use auto mode while having allow rules in Monitor, the number of confirmation prompts may increase, so it's worth being aware of this behavioral change.

New output style "Concise" to omit preambles and running commentary (v2.1.237)

A built-in output style called "Concise" has been added. Claude states results first and omits preambles and running commentary, while performing the work itself with the same level of care. Select it from Output style in /config.

For those who frequently review response logs after the fact, this feels like an additional option that makes it easier to follow the key points without preambles or commentary.

https://x.com/ClaudeDevs/status/2090245922685063634

Fix for prompt caching when using an LLM gateway (v2.1.237)

A bug has been fixed where prompt caching was not working for sessions using an LLM gateway or custom base URL.

For those using Claude Code via Bedrock or an internal gateway, this is a fix that should impact both cost and response speed.

Target Versions and Dates

New Features

• The environment variable ANTHROPIC_DEFAULT_MODEL has been added. It specifies the model that new sessions start with. (v2.1.236)
• notify_when_idle has been added to cross-session SendMessage. It allows you to request that another session on the same machine send a single notification the next time it becomes idle (macOS / Linux). (v2.1.236)
• [VSCode] Screen reader support for transcripts has been added. Live announcements for responses, permission requests, errors, and state changes, as well as heading navigation per turn, are now available. (v2.1.236)
• A built-in output style "Concise" has been added. It states results first and omits preambles and running commentary, while performing the work itself with the same level of care. Select it from Output style in /config. (v2.1.237)

Security

• In the macOS sandbox, wildcard read-deny rules (e.g., **/.env) now take priority even inside read-allowed areas. Directory contents that match are also covered, and it is no longer possible to bypass this by renaming the targeted files. (v2.1.236)
• A bug has been fixed where the managed settings approval prompt sometimes did not appear at startup, while the first keypress was being accepted as approval. (v2.1.236)
• While auto mode is enabled, Monitor allow rules are temporarily disabled, and Monitor commands are now confirmed in the same way as Bash commands. (v2.1.236)

Improvements

• In auto mode on Bedrock, Vertex AI, and Foundry, as well as when telemetry is disabled, the classifier now uses the same defaults as on the Claude API (including classification with severity scores). (v2.1.236)
• The auto mode git status check can no longer be tricked into misidentifying the tree as clean by the repository's status.showUntrackedFiles=no setting. (v2.1.236)
• If a slash command is mistyped or a command unavailable in the current session is entered and Enter is pressed, it now reports that fact rather than executing the closest fuzzy match. Prefix matches and aliases continue to work as before. (v2.1.236)
• SIGTERM in print / SDK mode no longer records interrupted turns or synthesized tool rejections before exiting. Running commands are still terminated, and the process exit code remains 143 as before. (v2.1.236)
• For /goal, idle sessions with a goal paused behind long-running background work will now automatically check in after 30 minutes (then 1 hour, then 2 hours) without waiting for the user to return. (v2.1.236)
• /usage now displays a credit consumption row for Team and Enterprise members. It also shows the limit row at 0% before any consumption has occurred. (v2.1.236)
• The /model picker now highlights only the name of the most recent model. (v2.1.236)
• Remote Control now marks a session as offline within a few seconds when the CLI exits or the terminal is closed. (v2.1.236)
• Startup performance has been improved, with session counter writes now happening in the background. (v2.1.236)
• The session title chip on the prompt border has been aligned to the right edge of the footer. (v2.1.236)
• Right-aligned footer items (goal indicator, session state, background agent state) and truncated notifications now align their right margin with other elements in the prompt area. (v2.1.236)

Fixes (Key Items)

Here are selected fixes related to stability and usability.

• Fix for prompt caching issue when using an LLM gateway: A bug has been fixed where prompt caching was not working for sessions using an LLM gateway or custom base URL. (v2.1.237)
• Fix for multiple features stopping after a directory is deleted: A bug (since v2.1.229) where clipboard copying, background housekeeping, background sessions, and local MCP logging would stop working has been fixed. (v2.1.236)
• Fix for fullscreen renderer permanently failing to start: A bug where a single startup failure would cause all subsequent attempts to fail has been fixed, so it now falls back to the legacy renderer rather than exiting on every startup. (v2.1.236)
• Fix for subprocess launch failure in WSL: A bug where an unhandled Promise rejection would occur when a subprocess failed to launch has been fixed (e.g., powershell.exe on WSL with Windows interop disabled; a regression from v2.1.234). (v2.1.236)
• Fix for skill hot-reload errors in SDK / VS Code: A bug (since v2.1.229) where an error would occur on every skill change after deleting the working directory has been fixed. (v2.1.236)
• Fix for SendMessage falsely reporting messages as sent: When a large number of messages are sent in a short time and exceed the capacity of a session's inbox, it now proactively rejects subsequent messages rather than reporting them as sent while actually discarding them. (v2.1.236)
• Fix for /model picker rendering overflow: A bug where the picker would render taller than the terminal has been fixed, so it now displays only as much as fits in the window with the rest accessible by scrolling. (v2.1.236)
• A quietly appreciated fix: A bug where the mascot Clawd's eyes and legs were rendered unevenly at certain font sizes in iTerm2 has been fixed (v2.1.236). It's a small detail, but since it's something you see every time you start up, having it aligned is nice.
• In addition, numerous other minor bugs have been fixed, including fullscreen mode re-rendering, tab titles in tmux, automatic model selection in Remote Control, session release for self-hosted runners, the character limit for session recaps, spinner tips, and error display in the cloud environment list.

Trying Out ANTHROPIC_DEFAULT_MODEL

Let's verify the difference in behavior between ANTHROPIC_DEFAULT_MODEL and the existing ANTHROPIC_MODEL. The JSON output from non-interactive mode (claude -p) includes a modelUsage field, where the model actually used in that turn appears as a key. We compared this with and without the environment variables.

$ claude --version
2.1.237 (Claude Code)

We run a command line that executes a prompt and retrieves modelUsage to verify behavior.

1. baseline (env var unset)

% unset ANTHROPIC_DEFAULT_MODEL
% unset ANTHROPIC_MODEL

% claude -p "Reply with exactly: ok" --max-turns 1 --output-format json | jq -r ".modelUsage | keys[]"

claude-opus-5[1m]

2. ANTHROPIC_DEFAULT_MODEL=sonnet

% unset ANTHROPIC_DEFAULT_MODEL
% unset ANTHROPIC_MODEL

% ANTHROPIC_DEFAULT_MODEL=sonnet claude -p "Reply with exactly: ok" --max-turns 1 --output-format json | jq -r ".modelUsage | keys[]"

claude-sonnet-5

3. ANTHROPIC_DEFAULT_MODEL=haiku

% unset ANTHROPIC_DEFAULT_MODEL
% unset ANTHROPIC_MODEL

% ANTHROPIC_DEFAULT_MODEL=haiku claude -p "Reply with exactly: ok" --max-turns 1 --output-format json | jq -r ".modelUsage | keys[]"

claude-opus-5[1m]

4. ANTHROPIC_MODEL=haiku

% unset ANTHROPIC_DEFAULT_MODEL
% unset ANTHROPIC_MODEL

% ANTHROPIC_MODEL=haiku claude -p "Reply with exactly: ok" --max-turns 1 --output-format json | jq -r ".modelUsage | keys[]"

claude-haiku-4-5-20251001

With ANTHROPIC_DEFAULT_MODEL=sonnet, we confirmed that the starting model changed to claude-sonnet-5, verifying that the variable is working. On the other hand, with ANTHROPIC_DEFAULT_MODEL=haiku, there was no change. According to the official documentation, if default / inherit / opusplan / haiku is specified, Claude Code ignores this variable. In the fourth example, where the same haiku was passed to ANTHROPIC_MODEL, claude-haiku-4-5-20251001 was used, showing that the two variables accept different values.

The two similarly named environment variables differ in both the values they accept and their priority, so I recommend testing them locally before adopting either.

Closing Thoughts

The two versions v2.1.236 and v2.1.237 feel like stability-focused releases where the number of fixes stands out more than the new features. In particular, the prompt cache fix when going through an LLM gateway has a large impact since it directly affects both cost and response speed.

If you are using Claude Code via a gateway, or using the sandbox on macOS, these are versions particularly worth checking. If any of the changes catch your eye, update and try them out.

References

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

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

https://code.claude.com/docs/en/output-styles

https://dev.classmethod.jp/articles/20260819-cc-updates-v2-1-235/