Start lower than you used to

The guide's main instruction is to recalibrate rather than migrate. Effort is the control for how much Opus 5.5 thinks, because thinking is always on, and the level names do not mean the same amount of thinking across the two models. A team that ran Opus 5 at High and copies the setting across can expect longer turns and more output tokens; whether that additional work is worthwhile is what the guide tells them to evaluate. The guide tells developers to set the level explicitly, start at Medium, and test several levels against their own evaluations.

That advice is the honest part of the document, and it sits awkwardly next to the headline claim. The comparison is scoped to Anthropic's coding and knowledge-work evaluations; it cannot settle the best setting for any other workload, which is precisely why the guide insists on the sweep. Anthropic is asking for trust in a comparison it has not published, while also telling readers not to rely on carried-over settings. Both can be true. Only one of them can be checked from the page.

What to delete from old prompts

Two kinds of legacy instruction come out. The first is wording that stood in for thinking: prompts that ask the model to write its reasoning into the reply. On Opus 5.5 that pattern can now be declined under a reasoning-extraction refusal category; the guide says to read summarized thinking blocks instead. The second is the consumer version of the same habit, from Anthropic's tips for chat users: "think hard" and similar phrases are ordinary prompt text rather than a control. The guide's own framing for developers is that lowering the effort level reduces thinking, and with it cost and latency, more reliably than prompt instructions do.

The guide also warns about a quieter cost. Thinking tokens count toward max_tokens even when the thinking itself is not returned, so a limit sized for Opus 5 with thinking off can cut replies short. For long agentic coding turns, Anthropic says a max_tokens of 128,000, the model's maximum, has worked well in its testing.

The part worth acting on

Strip the unverifiable comparison away and the operational advice still stands on its own. Set effort explicitly instead of inheriting it. Reserve xhigh and max for work where a quality gain has been measured. Change effort per message (in beta) rather than at the top level if the prompt cache matters, since a top-level change invalidates it. None of that depends on whether Medium equals High. It depends only on the defaults having moved, which is documented and checkable in any account.

Sources

  1. [1] Anthropic, "Prompting Claude Opus 5.5," Claude Platform DocsRead source
  2. [2] Anthropic, "What's new in Claude Opus 5.5," Claude Platform DocsRead source
  3. [3] Anthropic, "Effort," Claude Platform DocsRead source
  4. [4] The Decoder, "Anthropic says outdated prompting habits are slowing down Claude Opus 5.5," November 14, 2026Read source