Anthropic Outlines New Prompting Best Practices for Claude Opus 5.5
Anthropic urges users to drop outdated habits like 'think hard' directives, advising single-pass prompting and optimized effort settings for Claude Opus 5.5.
Anthropic has issued new guidance for users of Claude Opus 5.5, cautioning that legacy prompting techniques developed for older artificial intelligence models can actively hinder performance. Currently serving as the default model across Anthropic's Pro, Max, Ultra, and Team subscriptions, Opus 5.5 operates differently from its predecessors. According to the company, many users are unknowingly slowing down response times by relying on outdated instructions that are no longer necessary for modern adaptive reasoning architectures.
At the top of Anthropic's list of practices to abandon is the inclusion of directive phrases such as "think hard" or "think carefully." About a year and a half ago, these phrases were commonly used to steer router behavior or encourage deeper cognitive processing. However, Anthropic explains that Opus 5.5 autonomously calculates its reasoning depth based on the specific request and selected effort setting. In internal evaluations, prompts that omitted these trigger phrases yielded faster initial response times without any measurable compromise in output quality. Furthermore, the model is built to handle complete multi-step tasks from a single prompt, meaning users no longer need to guide it through incremental chat turns or request preliminary drafts before asking for a final document or spreadsheet.
Key prompting adjustments for Opus 5.5
Anthropic highlighted several specific operational adjustments to maximize the capabilities of Opus 5.5. Rather than manually transcribing structured data from charts or tables, users should upload visual assets or data files directly. The vision capabilities in Opus 5.5 have seen substantial improvements; internal benchmark data indicates that the model at its lowest reasoning effort reads values from dense charts with higher accuracy than Opus 5 achieved at its highest effort setting.
When providing instructions for design tasks, vague commands such as prohibiting "generic design" often result in the system simply swapping one default look for another. Anthropic recommends explicitly identifying unwanted design elements—such as pill-shaped buttons or cream-colored backgrounds—or providing an established house style guide. Additionally, connecting Opus 5.5 directly to external tools, files, and internal systems using connectors allows the model to retrieve context autonomously, leading to significantly more tailored results.
Cost optimization and model selection
The default operational setting for Opus 5.5 is set to "Medium" effort, accessible via the menu option located beside the send button. Anthropic’s internal testing demonstrates that Opus 5.5 configured at Medium effort matches the benchmark performance of Opus 5 operating at "High" effort. Because higher effort tiers increase token consumption without guaranteeing superior outputs, Anthropic advises escalating to "High" effort only when Medium proves insufficient for a given task.
This cost dynamic also impacts how teams choose between model tiers. Previously, organizations frequently relied on Sonnet for routine work while reserving Opus models for complex projects. However, Anthropic notes that running Sonnet at "Extra High" or "Max" effort settings creates expenses comparable to utilizing Opus 5.5 at standard levels. Because a more capable model can complete multi-part assignments in fewer total steps, the higher per-token price of Opus 5.5 can be partially offset by its operational efficiency.
What it means for developers
For software engineers and system architects, Anthropic's recommendations require a shift in how automated workflows and system prompts are constructed. API integrations built around sequential, conversational prompt chains should be refactored into comprehensive single-message prompts that clearly specify the ultimate goal, desired formatting, and criteria for when the model should seek clarification. Omitting obsolete phrases like "think hard" from automated prompt templates can directly reduce application latency.
Furthermore, developers building backend workflows can leverage direct file attachments and system connectors rather than building pre-processing pipelines to extract text from images or tables. Managing these evolving model capabilities and testing performance across different providers can be streamlined through unified access platforms. Developers looking to experiment with and deploy these advanced models can access Claude, GPT, Gemini, Grok, DeepSeek, and other top AI models cheaply through one API at https://apixoai.online. By aligning system prompts with Anthropic's updated standards, engineering teams can optimize token usage, minimize latency, and build more effective AI-driven software.
Source: Anthropic says outdated prompting habits are slowing down Claude Opus 5.5 — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
