Getting the most out of Opus 5.5 in Claude and Claude Code
Points and comments are a snapshot, not live.
Opus 5.5 requires less prompting and handles longer autonomous multi-step coding tasks better.
The guide covers how to work with Opus 5.5 in Claude apps and Claude Code. Key changes: it always thinks before replying, so "think carefully" prompts should be removed; it works longer on its own, so users should define "done" and hand over the whole task; it communicates what it did and needs plainly. For long runs, the guide recommends a CLAUDE.md rule telling it when to stop and when to keep going, splitting work across subagents for large audits, and keeping a task list in a file to survive context window summarization. It also notes design work benefits from listing specific styles to avoid, and flagged messages switch to an older model with a notice.
What commenters are saying
Commenters widely praised Opus 5.5, with one user reporting 12 PRs reducing CI time from ~10 to 4 minutes and billing minutes down 60%. Another noted it works great when goals are well-defined but requires more steering for hill-climbing tasks. The thread split over value: some canceled Anthropic subscriptions, saying Chinese models produce better code, while others defended Opus 5.5 as cheap and capable. Several debated whether Fable remains useful, with some finding Fable better for reviews and high-level planning despite lower benchmark scores. One user advised still knowing architecture yourself rather than relying solely on goals.