Does ultrathink permanently change my effort setting?
No. Classmethod's test showed the turn with the word ran at xhigh and the next turn went back to medium, with the saved setting untouched.
Only typing the word again within the same turn raises the rest of that turn.
I use Bedrock or Vertex. Am I affected?
Per both articles, the change is limited to the interactive terminal using the Anthropic API directly, and Bedrock, Vertex and Foundry are not covered.
In those environments, adjust effort with /effort instead.
Can I use CLAUDE_CODE_DISABLE_ULTRATHINK?
A changelog digest page says v2.1.316 added it to turn the keyword off. I found no official documentation or test, so try it once in your environment before rolling it out.
Another route is setting CLAUDE_CODE_EFFORT_LEVEL or maxEffortLevel below xhigh, where the lower cap wins.
How much more does xhigh cost than medium?
Neither article gives Token counts or dollar figures, and I have no reliable number. All that is certain is that reasoning tokens on that turn increase.
To learn your actual difference, compare before and after with OpenTelemetry or turn timings.
Claude Code v2.1.311 (October 26, 2026) changed a small habit many people have used for a long time: adding ultrathink to a prompt. It used to add an instruction to the prompt without changing the effort level sent to the API; now, in the interactive terminal, a turn that includes the word runs at xhigh effort. What follows rests mainly on the changelog's description and two third-party articles' tests; I did not read the official documentation directly.
Per the changelog, on the Anthropic API an ultrathink turn runs at xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1. Classmethod notes the default for Opus 5.5, Sonnet 5.5 and Haiku 5.5 is medium, so this is a clear jump. The word shows a rainbow highlight only where it can actually raise effort; it does not trigger inside quotes or file paths, nor in a prompt starting with /. Typing ultrathink while Claude is mid-work raises the rest of that turn.
Classmethod checked with $CLAUDE_EFFORT: a turn with ultrathink reported xhigh and the next turn went back to medium, so it affects a single turn and your saved effort setting is untouched. They also queued a message with ultrathink mid-turn, and later Bash calls became xhigh while earlier ones stayed medium, though that was one run. In non-interactive claude -p, the value stayed medium even with the word. Both articles caution that the change is limited to the interactive terminal using the Anthropic API directly; Bedrock, Vertex and Foundry are not covered, and there you set effort with /effort.
v2.1.316 added the environment variable CLAUDE_CODE_DISABLE_ULTRATHINK to turn the keyword off; I took that from a changelog digest page and found no documentation or test of it. Caps also override the keyword: CLAUDE_CODE_EFFORT_LEVEL set below xhigh, a maxEffortLevel below xhigh, or an organization limit, with the lower cap winning. What ultrathink actually does when a cap is below xhigh, Classmethod says it did not test and the changelog does not say.
Neither article gives Token counts or dollar figures, so I cannot give you a number for how much more xhigh costs than medium, only that reasoning tokens on that turn increase. Opus 4.6 and Sonnet 4.6 have no xhigh level and their behavior is unknown. The -p result comes from one author's test.
If ultrathink is already baked into your prompt templates, skills or colleagues' habits, those turns now run at xhigh, and bills and wait times may rise without any notice. Three steps: grep your templates and CLAUDE.md for the word, compare before-and-after with OpenTelemetry or turn timings, then decide whether to add it by hand only for genuinely hard tasks or switch it off entirely with a cap or the environment variable. Until you have measured, do not assume it is worth it.