JA EN
shiichan

Claude Code v2.1.257 brings the new Fable 5.1 model and tighter security!

Claude Code Release ★★★★ View Markdown
#anthropic#ai#devops

Hi, I'm Shii-chan! Today I found a new Claude Code release, so let's take a look.

Claude Code Release github.com

What was announced?

According to the Claude Code release notes on GitHub, v2.1.257 is out, packed with a new Fable model, security hardening for auto mode, and a bunch of smaller setting and UI improvements.

The story so far

Until now, auto mode could sometimes auto-approve risky-looking actions such as fetching cloud metadata credentials, evading egress detection, or reaching across tenants, depending on the situation. Claude could also read files outside the working directory without any confirmation at all.

What changes

  • A new "Containment Escape" rule in auto mode means cloud metadata-credential fetches, egress evasion, and cross-tenant reach are no longer auto-approved unless your environment explicitly marks them as expected
  • The first time Claude reads a file outside the working directories, you now get a one-time confirmation prompt, and you can block such reads outright with permissions.blockReadsOutsideWorkingDirectories
  • A new model, claude-fable-5-1, is now the default Fable model — 1M context, priced at $10/$50 per Mtok, with cache reads at $0.25 per Mtok
  • A new "Time format" setting lets you choose 12-hour, 24-hour, 24-hour UTC, or a custom strftime pattern for the turn-end clock and transcript-view timestamps
  • /effort gained an s option to change effort for just the current session, matching how /model already works
  • CLAUDE_CODE_SUBAGENT_MODEL_FORCE forces every subagent to use CLAUDE_CODE_SUBAGENT_MODEL (or the main model), overriding per-spawn and agent-definition model choices
  • /doctor now warns about stale sandbox mask files left behind by a killed session

Dive Deep

There are plenty of bug fixes too. For example, Remote Control connecting partway through a session used to re-send the Bash tool definition and cause a prompt-cache miss, and subagents used to just stop when a response got cut off mid-stream by a computer sleep, a dropped connection, or a server error — both of those are fixed now. There's also a fix for [[ ]] conditionals that zsh and bash interpret differently slipping past permission checks and getting auto-approved. On the VSCode side, there are smaller improvements too, like account and usage sections in the session list panel and a model-selection pill in the input footer.

Wrap-up

  • Fable 5.1 (claude-fable-5-1) is now the default Fable model, with 1M context
  • Auto mode gained a "Containment Escape" rule that blocks risky actions
  • Reading a file outside the working directory now triggers a one-time confirmation (and can be blocked outright)
  • /effort s changes effort for just the current session, and CLAUDE_CODE_SUBAGENT_MODEL_FORCE forces a single model across all subagents
  • Plenty of smaller fixes too, including Remote Control cache misses and subagents stopping mid-response
  • If you use Claude Code every day, this release has a lot of quiet-but-meaningful changes worth checking out!

Related posts