Anthropic makes Claude Opus 5.5 the default knowledge-work model and retires old prompting habits

Anthropic has published a knowledge-work guide that makes Claude Opus 5.5 the recommended default model for everyday work across its paid plans, paired with a list of prompting habits the company says users should drop. According to a dev.to writeup, the guide appeared on November 12 based on web-archive records and was surfaced by The Decoder on November 14. It positions Opus 5.5 as the go-to model in Claude for research, analysis and finished documents, decks and spreadsheets on Pro, Max, Ultra and Team plans.

Reasoning moves to the model

The headline change concerns instructions rather than the model itself. According to Anthropic, thinking is always active in Opus 5.5 and the model decides on its own how deeply to reason about a request. That makes prompts such as "think hard" or "think carefully" redundant — in Anthropic's own testing, removing them meant replies started sooner with no clear loss in quality.

The platform documentation adds a hard constraint: adaptive thinking on Opus 5.5 cannot be switched off, and requests that try receive an error. The only user-side lever is the effort setting, and Anthropic recommends beginning at Medium, which it calls the new default. In the company's coding and knowledge-work evaluations, Opus 5.5 at Medium reportedly matched or beat Opus 5 running at High effort, while higher settings reason for longer, cost more and do not consistently produce better work.

Seven habits to update

The guide lists seven practices it wants users to change:

1. Connect Claude to work files and systems instead of pasting content in.
2. Describe the finished result in the very first message.
3. Delete reasoning-coaxing lines from saved prompts.
4. Stay at Medium effort by default and escalate only when a task demands it.
5. Attach charts rather than retyping them as text.
6. Name the specific design styles to avoid instead of asking for something that does not look generic.
7. Share a house style so output follows organisational conventions.

Pricing and vendor-reported benchmarks

The same page lists API pricing. Opus 5.5 costs 4 dollars per million input tokens and 20 dollars per million output tokens, sitting between Fable 5.2 at 8 and 40 dollars and Sonnet 5.5 at 2 and 10 dollars, with Haiku 5.5 at 0.10 and 0.50 dollars.

The guide also cites internal tests: 16 of 18 generated reports passed an automated check covering every figure and quote; a fictional merger analysis finished in 63 minutes versus 93 minutes for Opus 5, at half the cost; and Factory, an early tester, reported results matching Opus 5 at High effort while consuming 20 to 25 percent fewer output tokens. All of these figures come from Anthropic and the testers it cites.

Why it matters

For teams running Claude in production, the advice is concrete and cheap to test: audit saved prompts and strip out reasoning instructions, keep effort at Medium for routine work and raise it per task, and stop transcribing charts into text. The deeper shift is architectural — reasoning depth is now a model-side control, so prompt engineering moves away from coaxing a model to think and toward describing outcomes and supplying context. Vendor guidance about a vendor's own model is self-reported by definition, but the claim that Medium effort matches High is easy to replicate on your own tasks before rewiring any workflows.