Anthropic's Opus 5.5 guide: start at medium effort, drop "think hard", and check how your agent loop reads a text-only turn
Anthropic's new Opus 5.5 prompting guide reads like a migration checklist for anyone running Claude inside automations.
What happened
Anthropic made Claude Opus 5.5 the default model on its Pro, Max, Ultra and Team plans and published a guide on prompting it. The Decoder summarized the chat side: drop phrases like "think hard" or "think carefully", because the model always reasons and scales that reasoning to the task and the effort level. In Anthropic's tests, responses without those phrases started faster with no measurable drop in quality.
The developer guide goes further. Effort is now the main control, and the default is medium, where Opus 5 defaulted to high. In Anthropic's testing, Opus 5.5 at medium matched or beat Opus 5 at high on coding and knowledge work evaluations. At a given level it also thinks more per turn than Opus 5, so a team that keeps its old setting should expect longer turns and more output tokens. Thinking counts toward max_tokens even when it isn't returned, and an integration that sends thinking disabled now gets a 400 error.
For unattended agents, the guide warns that some progress updates end the turn with text instead of a tool call. A loop that treats that as the end of the task stops there. Anthropic suggests keeping a checklist the model updates, sending a short message naming open items, and stopping after two or three automatic continuations so a stuck run can be reviewed.
My take
Many Claude integrations in business automations sit inside n8n workflows or small services that were written before these behaviors existed. Before switching a model ID I'd check four things: remove any thinking disabled parameter, set effort explicitly instead of relying on the default, raise max_tokens so thinking doesn't truncate replies, and look at how the loop handles stop_reason end_turn.
That last one matters most for unattended work. If the loop exits on the first text-only turn, a job can report itself finished with half the records untouched. A completion check that lives outside the model, plus a hard cap on continuations, keeps that failure visible.
The guide also notes that changing top-level effort between requests invalidates the prompt cache, with a per-message effort option in beta. The guide doesn't give cost numbers for any of this, so run your own evals at low and medium before you move production traffic.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
