Claude Code 2.1.311: model settings load from a signed file, ultrathink runs at xhigh

what changed

Claude Code 2.1.311 was listed on 26 October 2026. These are the changes that matter most for teams:

• Model settings on Bedrock, Mantle, Vertex and Foundry now load from a signed file on downloads.claude.ai. Setting CLAUDE_CODE_MODEL_CATALOG=off turns this off.
• In the terminal, ultrathink now runs its turn at xhigh effort on several models: Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5 and Claude Fable 5.1, through Anthropic’s API.
• JSON and stream-json output now escapes DEL, C1 control and Unicode format characters as \uXXXX. Parsed values stay the same.
• DISABLE_COMPACT in settings files and in mod $.env.set can now only turn compaction off.
• self-hosted-runner gained --hold-idle-file and --max-hold-min to keep an idle session while the host still reports running processes.

why it matters

The model catalog change adds a network fetch at startup for teams that run Claude Code against a cloud provider. If your developer machines or CI runners sit behind a strict egress allow-list, that fetch may be blocked. It also means model settings can change without a Claude Code upgrade.

The ultrathink change is about cost and latency. A turn that used to run at a lower effort now runs at xhigh, so people who use the keyword often may see slower and more expensive turns.

The output escaping change matters to anyone who parses Claude Code output with scripts that match raw characters.

what we would do

• Check whether downloads.claude.ai is reachable from your runners and laptops. Decide on purpose whether to allow it or set CLAUDE_CODE_MODEL_CATALOG=off.
• Tell the team that ultrathink now costs more per turn, and watch usage for a week.
• Run your output-parsing scripts against 2.1.311 before you roll the version out to CI.
• Pin the Claude Code version in shared images, so the change reaches everyone on the same day. Our enable service covers rollout like this.