Anthropic Releases Prompting Guide for Claude Opus 5.5 with Effort Calibration and Agentic Updates

Anthropic has released a comprehensive prompting guide for Claude Opus 5.5, highlighting significant changes in effort calibration, thinking behavior, and agentic workflows compared to Claude Opus 5. The model generates output tokens more than 30% faster than its predecessor and often completes tasks with fewer tokens. In agentic coding and code review, the default medium effort level matches or exceeds the performance of Opus 5’s high effort, while maintaining better persistence in multi-hour autonomous tasks.

Effort and Thinking Calibration

Effort is the primary control for thinking volume in Opus 5.5, with Thinking always enabled. Unlike Opus 5, which allowed thinking: {"type": "disabled"}, Opus 5.5 does not support disabling thinking. Developers must migrate prompts by starting at medium effort (the default) and measuring latency/quality, or starting at low effort if optimizing for speed, removing any instructions that forced reasoning into the response text. Instead, use display: "summarized" to view reasoning in thinking blocks.

• Default Effort: Changed from high (Opus 5) to medium (Opus 5.5).
• Max Tokens: Set to 128,000 for agentic coding to accommodate hidden thinking tokens.
• Cache Impact: Changing the top-level effort value invalidates the prompt cache; per-message changes require the Beta Per-message effort change feature.

Agentic Workflows and Progress Updates

Opus 5.5 may end turns with text-only updates (stop_reason: "end_turn") during long tasks, which can stall unattended loops. To handle this, treat text-only turns as reports rather than completion, use checklists to track open items, and limit automatic continuations to 2–3 attempts.

• Progress Updates: Short updates between tool calls are returned as thinking blocks, not text blocks. To display these, set display: "updates" (Beta header thinking-display-updates-2026-08-18).
• Verbatim Content: Declare a simple tool for user messages to pass code snippets or exact text verbatim, as default thinking blocks may omit this.
• Time Signals: Adding elapsed time budgets (e.g., elapsed 340s / 1200s) helps parallelize multi-agent workflows, reducing completion time without significantly impacting quality.

Safety and Security Refusals

Opus 5.5 introduces new safety classifiers for biology, cybersecurity, and reasoning extraction.

• Reasoning Extraction: Requests forcing internal reasoning into response text are refused under the reasoning_extraction category. These refusals are returned directly to the user and do not trigger automatic fallback to other models.
• Cybersecurity: Vulnerability discovery in source code is allowed, but high-risk dual-use activities are prohibited.
• Biology: Same safeguards as Claude Fable 5.1; daily health/education questions are unaffected.

Visual Inputs and Prompt Injection Defense

Opus 5.5 interprets visual inputs (charts, diagrams) more accurately than Opus 5, even at lower effort levels. For dense technical drawings, high-resolution images and image processing tools (crop, zoom) improve accuracy.

• Prompt Injection: The model is more resistant to indirect prompt injection. To further secure user-pasted content, wrap it in unique tags (e.g., <pasted_content id="ab12">) and instruct the model to treat internal instructions as untrusted unless explicitly requested by the user.