Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-25 · high
Anthropic says lowering effort doesn't cut response length, only reasoning depth, and wants evals rerun from scratch.
Why now: Anthropic's effort docs now contradict the xhigh-for-coding advice teams adopted for Opus 4.7 and 4.8, leaving pinned evals unverified until rerun.
Anthropic reversed its coding guidance, telling developers to start Claude effort at high instead of xhigh, per its documentation. The stakes: effort controls how much Claude thinks, not how much it says. Per the docs, changing effort doesn't reliably shorten responses.
The new guidance frames low and medium as the main dials for cost and latency, not as fallback settings for when high runs too slow.
Anyone dropping effort to cut verbosity is paying for less reasoning and getting the same response length back, per the documentation. The fix Anthropic recommends is prompting for length directly instead of treating effort as a verbosity dial.
Effort also changes how Claude uses tools. At lower levels it combines operations into fewer calls and skips preamble before acting. At higher levels it makes more calls and explains its plan first.
Anthropic's own instruction is to run a fresh effort sweep on your evals rather than reuse old ones. Any production prompt still pinned to xhigh from the Opus 4.7 or 4.8 era should be treated as unverified until it's rerun against current behavior. The reversal undoes advice tied to those two releases, so anyone who set defaults then is running on guidance Anthropic no longer stands behind.
Each link below shares sources, entities, or timing with this story.
Anthropic released Claude / Shared entities / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Anthropic released Claude); both cover Anthropic, Effort; overlapping topics (call, effort).
Opus built by Anthropic / Shared entities / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Opus built by Anthropic); both cover Anthropic, Opus; overlapping topics (advice, behavior).
Linked by a graph relationship (Opus built by Anthropic); both cover Anthropic, Opus; overlapping topics (capability, coding).
Opus built by Anthropic / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Opus built by Anthropic); both cover Anthropic, Opus; overlapping topics (anyone, capability).
Linked by a graph relationship (Opus built by Anthropic); both cover Anthropic, Opus; overlapping topics (call, coding).
Opus built by Anthropic / Shared entities / Same source domain / Earlier coverage
Linked by a graph relationship (Opus built by Anthropic); both cover Anthropic, Opus; reported by the same outlet (platform.claude.com).
Opus built by Anthropic / Shared entities / Earlier coverage / Tension
Linked by a graph relationship (Opus built by Anthropic); both cover Anthropic, Opus; earlier Anthropic coverage from 2026-07-10.
Anthropic partners with White House / Shared entities / Earlier coverage / Tension
Linked by a graph relationship (Anthropic partners with White House); both cover Anthropic, Opus; earlier Anthropic coverage from 2026-06-17.