Alongside the September 22, 2026 launch of Claude Opus 5.5, Anthropic shipped a prompting guide that tells developers their old Opus 5 habits can now cost more and slow things down. The headline changes: the default effort dropped from high to medium, thinking can no longer be switched off, and lines like "think carefully" should come out of chat prompts because the model now sets its own thinking on every reply.
2026-09-28 · by Moe Ameen
Anthropic published a prompting guide specific to Claude Opus 5.5 as part of the model's September 22, 2026 launch, and the advice drew fresh attention this week from outlets including Search Engine Journal. The guide's framing is unusual for a model release: rather than promising better output from the same prompts, it warns that settings carried over from Claude Opus 5 can now run longer and cost more, and it walks through the behavioral changes that make old habits misfire. Anthropic notes that existing Opus 5 prompts should still perform well without changes — the guide is about tuning, not rescue.
The central change is effort. Opus 5.5's default effort level is `medium`, one step below Opus 5's `high` default, and thinking is now always on — the model no longer accepts requests with thinking disabled. Because Opus 5.5 tends to think more per turn at any given level, reusing an Opus 5 effort value can produce longer turns and more output tokens. Anthropic's recommendation is to set effort explicitly, start at `medium`, and test several levels against your own evaluations rather than assuming the level names carry the same meaning across models. In its testing, Opus 5.5 at `medium` matched or beat Opus 5 at `high` on coding and knowledge-work tasks, in fewer steps and with fewer tokens, and it advises sizing `max_tokens` high enough (up to the model's 128,000 maximum on long agentic turns) because thinking counts toward that limit even when the thinking text isn't returned.
For chat applications specifically, the guide says to remove system-prompt lines that tell Claude to "think carefully" before answering. The model decides for itself how much to think, effort is the real control, and in Anthropic's testing removing such a line made replies start sooner with no clear decline in quality. The rest of the guide covers agentic and multi-app patterns: an unattended agent should not treat a text-only end of turn as proof the task is finished; progress updates now arrive as "progress-update" thinking blocks whose text is empty at the default display setting (so a client that renders only text can look silent during a long turn); multi-agent harnesses can be sped up by feeding the model an elapsed-time budget; and text a user pasted from elsewhere should be wrapped in tagged blocks, because Opus 5.5 resists prompt-injection better than any earlier Opus model when it knows which text is the user's own. It also notes a new `reasoning_extraction` refusal category and advises naming specific styles to avoid for frontend work rather than asking it to "avoid a generic AI look." Verify any figure against Anthropic's own docs before building around it; the guidance is versioned to the model and can change.
Read this guide as a creator and the honest conclusion is that it's a to-do list you'd rather not own. Retesting effort levels, sizing token budgets, rewriting agent stop conditions, tagging pasted text against injection — that's the work of running a raw model well, and it's exactly the layer [Kompozy](/) already owns so you don't. Kompozy's Text Posts, Blog Articles, and Newsletters are generated by this class of Claude and OpenAI model, but the prompting is encoded in the engine: a [Persona Brief](/glossary/persona-brief) that fixes voice and banned words, per-format prompts tuned for each output, and model settings managed for you. When Anthropic changes a default or flips thinking to always-on, that's Kompozy's problem to absorb, not a prompt in your notes app to go re-tune.
And the guide can only ever make the words better — it can't make the post. That's the seam Kompozy fills. From one idea it generates the things a text model can't: talking-head [Persona Shorts](/glossary/persona-shorts) and HeyGen avatar video with a consistent face and voice, brand-exact Carousels and Quote Graphics rendered through [HyperFrames](/glossary/hyperframes), Photo Posts and Infographics — then [Autopilot](/glossary/autopilot) and a per-post review pipeline caption, reframe to 9:16, 1:1, and 16:9, schedule, and publish the set across the eight social platforms plus a blog and a Mailchimp newsletter. If you want the model detail, see [Claude Opus 5.5](/ai-tools/claude-opus-5-5); if you're deciding whether learning to prompt it is worth it versus an engine that ships content, the [Kompozy-vs-DIY-prompting comparison](/alternatives/claude-opus-5-5-prompting-guide) and the honest [prompting-guide review](/reviews/claude-opus-5-5-prompting-guide) lay out where each stops. The guide makes the brain sharper; Kompozy turns it into a published week.
Effort calibration. Opus 5.5 defaults to medium effort (Opus 5 defaulted to high) and thinking can't be disabled, so a setting carried over from Opus 5 can run longer and cost more. Anthropic advises setting effort explicitly, starting at medium, and testing several levels against your own evaluations, since the level names don't map to the same amount of thinking across models.
For Claude Opus 5.5 chat prompts, yes. The model decides how much to think on its own, and Anthropic's testing found that removing "think carefully" style lines made replies start sooner with no clear decline in quality. Effort, not a prompt instruction, is the intended control for how much the model thinks.
No. Unlike Opus 5, which accepted thinking disabled at high effort or below, Opus 5.5 does not — requests that try to disable thinking return an error. To get less thinking and lower latency, Anthropic says to lower the effort level (start at low and measure) rather than disable thinking.
Indirectly at best. The guide is engineering advice for building reliable apps and agents on the API — effort, token budgets, agent loops, injection defenses. It makes the model's text better, but it doesn't produce video, images, carousels, or scheduled posts. For finished, published content, you still need a generation-and-publishing engine like Kompozy on top of the model.