Claude Code 2.1.311: ultrathink now raises effort mid-turn

Claude Code 2.1.311, released on October 26, 2026, comes down to one change that touches a habit a lot of people already rely on, plus a handful of fixes that mostly matter in VS Code and in request reliability. The ultrathink keyword, which you already typed to ask for deeper reasoning on a turn, gains a new power: it now actually raises the turn's effort to xhigh, even when you type it while Claude is already working.

ultrathink now raises effort, even mid-turn

Until now, including ultrathink in a prompt only added an instruction asking for deeper reasoning on that one turn, without touching the effort level, as documented on the effort levels page. 2.1.311 gives it a real effect on effort, covered in the section on using ultrathink for one-off deep reasoning: typed in the terminal, the word runs its turn at xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 running through Anthropic's API, and typed while Claude is still working, it raises the rest of the current turn. On Amazon Bedrock, where the word can't raise a turn this way, you need /effort instead. When it works, the word shows up in rainbow colors in the terminal; if it stays plain text, the docs suggest retyping it outside a path, a command, or quotes, in a message that doesn't start with /, or checking that your CLAUDE_CODE_EFFORT_LEVEL or maxEffortLevel actually allows xhigh. In practice, if a task turns out harder than expected partway through, one word in your next message is enough to raise effort for whatever is left, without interrupting the turn.

The rest: VS Code, model names, and reliability

The VS Code extension finally adds syntax highlighting to code blocks in Claude's replies, across the light, dark, and high-contrast dark themes. Model names and IDs drop the [1m] suffix wherever it has no effect, though the suffix is still accepted as input; if any of your scripts match on that string in a model name, make sure they tolerate its absence. On Bedrock, Mantle, Vertex, and Foundry, Claude Code now loads model settings from a signed file hosted on downloads.claude.ai; CLAUDE_CODE_MODEL_CATALOG=off turns this off. Three fixes are worth knowing about. Claude Code was retrying requests the API had explicitly marked not to retry, which could hold a turn for minutes, or indefinitely under unattended retry mode. Subagents couldn't use MCP tools from a server that finished connecting after the subagent had already started, a blind spot that hit MCP servers slow to connect. And a standard-context session used to stop cold with a credits error the moment a request passed 200,000 tokens; Claude Code now compacts the conversation and keeps going instead of cutting the turn short. The rest of the 56 fixes are scattered (startup freezes, SubagentStop hooks, /mcp, /rewind, Remote Control, cloud sessions, the VS Code extension): nothing that changes a habit unless you're directly affected.