Anthropic published its prompting guide for Claude Opus 5.5 in the official Claude platform documentation on September 28, 2026. The guide doesn't talk about how strong the model is — it covers one thing only: which prompts and engineering setups carried over from Opus 5 need to change. There is a single core change — on Opus 5.5, thinking can no longer be turned off.
The biggest change: effort becomes the primary control knob
Opus 5 allowed disabling thinking at high effort and below; Opus 5.5 no longer accepts that option. Instead there is the effort setting: 5.5 defaults to medium while Opus 5 defaulted to high. In Anthropic's testing, 5.5 at medium already matches or beats Opus 5 at high on coding and knowledge-work tasks, with low coming close on some coding evaluations at much lower cost.
Don't carry your old effort value over when migrating: at the same level, 5.5 thinks more per turn, especially at xhigh and max, producing longer turns and more tokens. Anthropic recommends starting at medium and testing each level against your own evaluations rather than copying your Opus 5 configuration. Two more easy traps: thinking counts toward max_tokens, so a limit sized for Opus 5 with thinking off can cut replies short — for long agentic tasks, 128,000 has worked well; and changing the top-level effort value invalidates the prompt cache, so use the beta per-message effort change to adjust a single turn while keeping the cache.
Four changes when migrating from thinking-disabled setups
If your integration ran with thinking disabled, the guide lists four changes: start at low effort and measure, moving to medium if quality drops; remove instructions that asked the model to write out its reasoning in the response as a substitute for thinking, and read it from summarized thinking blocks instead; re-test the thinking-disabled mitigations from the Opus 5 era and delete whatever is no longer needed; read responses by block type instead of assuming the first block is text — a thinking block's content is empty under the default display setting.
There is also something new on the safety side: alongside biology and cybersecurity, there is a new reasoning_extraction refusal category — prompts that push the model to reproduce its internal reasoning in the response text get declined, returning a stop_reason of refusal.
Two traps for unattended agents
On long tasks, Opus 5.5 proactively reports progress, and some of those updates end the turn with plain text. Treating a text-only turn ending as task completion stops an unattended loop too early. The guide recommends treating it as a report, not proof of completion: track the task's parts in a checklist or file, and if items remain open with no blocker stated, send a short user message naming them; stop after two or three automatic continuations on the same task and review it manually.
The other trap is "silence": progress updates live inside thinking blocks, whose text is empty under the default display, so a client that renders only text blocks will perceive long turns as dead air. Turning on the beta display: "updates" delivers a summary of each update.
Opus 5.5 didn't add new parameters this time — it swapped in a different cost model: with thinking always on, token savings come not from switching thinking off but from calibrating effort to the level the task actually needs. Migration, at its core, isn't swapping a model name; it's re-verifying, one by one, every engineering assumption written for an era when thinking could be disabled.
