Claude Code’s ultrathink is back—and you can turn it up mid-task

There’s a word back in Claude Code: ultrathink.

Addy Osmani posted a reminder: if a task is harder than expected, type this word while Claude is already working, and the reasoning effort for the rest of the current turn will be increased. It’s in the terminal and available in the latest version.

First, some context. Claude Code is a coding agent that runs in the terminal. ultrathink isn’t a new model or another product; it’s more like a temporary switch: it makes the model think a little longer during this turn. It’s suited to tasks that seem simple when you start but reveal a deep pit halfway through.

Adam Wolff put it more directly: ultrathink is back in the CC CLI, meaning the Claude Code command-line interface. It had been turned off for a while on the grounds that the word was odd and could confuse people. But he likes odd, and he doesn’t think it’s that hard to understand. Most of the time he uses Opus 5.5 on medium; on certain turns, when he wants a little extra kick, he types ultrathink.

This description is pretty accurate. The product compresses the abstract parameter of raising reasoning effort into a word with a very low input cost. Odd as it is, it’s faster than flipping through a settings page.

Someone in the comments defined it as: an official command specifically for 'this was supposed to be a five-minute fix.' Others realized it only in hindsight, saying hard problems often don’t show their true face until ten minutes in. That’s where the value of mid-turn input lies: when you notice something’s off, you can still crank it up on the fly.

At this point, the only things confirmed are the latest terminal version and the current-turn boost. Whether desktop supports it, whether it affects subagents that are already running, and whether it touches the cache—the public replies in the source material still don’t give clear answers. Someone asked what level it raises medium to, and there’s no official explanation either. The mention of xhigh in the comments should be treated only as discussion, not as a spec sheet.

Regular users can understand it as letting the AI think a little longer on hard problems. Readers with a development background will care more about the boundaries: does it only affect the current turn, or can it propagate to subagents? Does changing the effort value reset the context cache? These questions matter more than a one-liner saying it’s useful.

One judgment is that ultrathink’s return isn’t just restoring an easter egg. It turns 'apply a temporary boost' into an executable action, preserving room for manual tuning without having to guess task difficulty at the start. Using an odd word as a switch is surprisingly straightforward.