Effort level: xhigh

The short version

This tier is newer capability, not a relabelled existing one — model generations that predate its introduction reject it outright rather than quietly running at their own ceiling instead. It's also the effort level Claude Code defaults to, which says something about how demanding the work it was built for typically is.

The one thing worth remembering here specifically

Requests at this tier need considerably more room reserved for the answer than a default-tier request does, because more of the shared budget goes to thinking before the visible response is written. A max_tokens ceiling copied over unchanged from a default-tier request is the most common way this tier produces a truncated answer with no error explaining why.

Where it doesn't earn its cost

Not every task that feels hard actually benefits from this tier — the gain is concentrated on genuinely open-ended, multi-step work, not on tasks that are merely unfamiliar or verbose. Treat a request for this tier as something to justify with a real, tested improvement over the default, not something to reach for by default once a task looks demanding.

Related

See when to use xhigh effort for the full decision framework — which tasks are a genuine fit, how to test before adopting it, and why the pull toward reaching for it more broadly is worth resisting — and max_tokens is too small once xhigh thinking is included for the specific failure mode above.