On July 24, 2026, Anthropic refreshed its dedicated prompt guide with the release of its new flagship model, the Claude Opus 5. With the dramatic improvement in the autonomous reasoning ability of the model itself, a new design philosophy has been introduced that fundamentally overturns traditional instruction methods.

The impact of the Opus 5, achieving top-tier intelligence at half the cost

Claude Opus 5, announced by Anthropic on July 24, 2026, has significantly reshaped the power landscape of the AI industry. The most notable point is that it achieves intelligence comparable to the top-tier Fable 5 model at just half the cost. The specific pricing remains unchanged from the previous generation Opus 4.8, at $5 per 1 million input tokens and $25 per output token. The benchmark results are also overwhelming, with Frontier-Bench v0.1, which measures practical software engineering capabilities, recording 43.3%, double the previous generation’s 21.1%, and far ahead of competitor GPT-5.6 Sol’s 34.4%.

Additionally, in ARC-AGI-3, which measures the ability to solve unknown problems, it achieved an astonishing score of 30.2%, replacing the 1.5% of Opus 4.8 and 7.8% of GPT-5.6 Sol. This performance improvement shifted the Opus 5’s position from a “model for special moments” to a “standard model for everyday use.” The official documentation also positions Opus 5 as the default model for Claude Max, and the revamp of the prompt guide focuses on how to efficiently unlock this overwhelming capability in everyday tasks. Please refer to the performance comparison chart below.

The revamped prompt guide sets new standards for AI use

The prompt guide, published exclusively for Opus 5, breaks many conventional wisdom about AI utilization. The most notable change is that model autonomy has been maximized to the extreme, eliminating the need for detailed human “micromanagement.” For example, instructions like “verify the content yourself after answering” and prompts like “generate subagents and share tasks,” which were essential in previous models to improve answer accuracy, are no longer necessary in Opus 5.

This is because the Opus 5 comes standard with the ability to autonomously verify tasks and coordinate between agents within the model. Conversely, leaving redundant verification instructions in prompts like traditional ones risks excessive self-validation, leading to unnecessary token costs and increased execution times. The new guide encourages users to focus on clearly stating the completion conditions of “what they need to achieve,” and recommends a “director-like” approach that leaves the process to AI’s autonomous judgment. Please refer to the following image diagram to see changes in the model’s autonomous behavior.

New parameter settings and design philosophies to control thinking

Five-level effort settings to balance intelligence and budget

The most practical new feature in Opus 5’s prompt guide is the extension of the effort setting, which specifies the “effort” of inference. This parameter, introduced in Opus 4.6, was organized into five levels in Opus 5: low, medium, high, xhigh, and max. This setting is not just about turning intelligence on or off, but about determining how much of the AI will allocate to the task. For example, for simple email summaries, low costs and time are reduced, while for complex system designs or unknown bug fixes, deep inference can be performed using xhigh or max, making it possible to distinguish between usage.

According to benchmark results, even with the same Opus 5 model, its performance curve can fluctuate significantly depending on Effort settings. For difficult problems like ARC-AGI-3, there have been reports of a 30.2% correct answer rate at a high setting with an evaluation cost of about $20,000, showing that the yield on budget far exceeds that of competing models. Users are expected to design prompts from the perspective of return on investment, focusing on “how high quality can be extracted within this budget and time,” rather than a fixed view of “how many points this model has to offer.” The intelligence curve for each effort setting is shown in the graph below.

Transparency brought by enabling default thinking mode

Another major feature of Opus 5 is that the “Thinking Mode,” which visualizes the deep reasoning process, is enabled by default. This mode allows users to check the logical steps Claude takes and the assumptions placed before giving a final answer, as well as the “behind” of their thinking, on the terminal or chat interface. This is not just a display gimmick; it aims to prevent AI from becoming a black box and to realize an auditable thought process.

According to the guide, Opus 5 flags confidence levels (low, medium, high) during thinking and continuously verifies internally whether its reasoning is correct. If a contradiction is found in the thought process, it will adjust its course on its own, leading to more refined answers. This increased transparency is driving adoption in highly reliable areas such as legal, financial, and cybersecurity. Additionally, by observing this inference process itself, users can easily debug to identify which parts of the prompt are misleading the AI.

How to Unlock ‘Extreme Abilities’ as Seen in the Leaked Internal Playbook

The power of context injection to achieve a 41% improvement in quality

The contents of Anthropic’s internal prompt playbook, leaked in May 2026 by a former researcher, deeply resonated with the Opus 5 guide and have renewed attention within the developer community. Among these ten tactics, the most powerful is “The Context Dump.” Rather than simply asking questions, it comprehensively outlines user expertise, project background, past methods, and current stagnation points. Internal tests have reported that this single change improved output quality by 41%.

The leaked playbook highlights Claude’s fundamental design philosophy. By default, Claude tends to give indirect and verbose answers out of a ‘sincere and helpful nature,’ but by setting appropriate context and assigning roles, it is said that this limiter can be lifted and sharp analytical capabilities unleashed. The new guide also suggests that this “context thickness” forms the boundary between general users who can only utilize about 30–40% of Opus 5’s potential and professionals who fully realize it. An example of the leaked prompt structure is summarized in the diagram below.

Brushing Up Thinking Through Self-Friendly Red Team Exercises

The fifth technique in the leaked playbook, “Red Team Yourself,” is truly effective when combined with Opus 5’s advanced reasoning capabilities. This is a method in which AI orders users to “mercilessly destroy any ideas or plans they propose.” The model plays the role of pointing out false assumptions, overlooked risks, and secondary impacts without bias. Anthropic’s product team itself always uses this method when making important decisions, thoroughly eliminating blind spots.

Opus 5 is more compliant with the “Constitution” than previous models and tends to faithfully adhere to the prescribed format, enabling more detailed and constructive critiques in this Red Team exercise. Furthermore, by combining “Pre-Mortem,” which analyzes the causes of project failures, the probability of discovering critical flaws before real-world operation has dramatically increased. Prompt design that deliberately positions AI not merely as a consenter but as a hostile verifier is the advanced utilization of the Opus 5 era.

Practical Application and the Boundary of Safety: Future Outlook

On-site Introduction and Challenges as Autonomous Agents

The ultimate goal of the Opus 5 guide is to elevate AI from a “chatbot that gives one-shot answers” to an “autonomous agent” that continuously interacts with its environment. In fact, for tasks that require operating terminals for long periods, repeatedly editing files and executing commands, Opus 5 delivers more than twice the performance of Opus 4.8. AutomationBench, which measures multi-step business automation, also recorded scores far ahead of the second-place group, demonstrating that the ability to autonomously complete everything from investigation to implementation and verification has reached a practical level.

However, looking at the absolute value, the completion rate of business workflows remains at only 26.0%, and it has not yet reached the stage where all tasks can be fully delegated to AI. Ultimately, humans must set appropriate completion conditions and guide AI thinking along prompt guides. Additionally, management costs in actual operations have emerged as new challenges, such as managing token consumption during long execution periods and countering unexpected side effects caused by autonomous behavior. The following diagram illustrates how agents autonomously perform tasks.

Turning Points in Design Philosophy in the Cyber and Biological Fields

When operating Opus 5, the most important thing users should be aware of is the safety safeguard restrictions. Anthropic, wary of the increased cyberattack capabilities as a byproduct of improved general intelligence in its models, intentionally adopted a hierarchical design that “has a higher ability to find vulnerabilities but less ability to write code that exploits them.” Specifically, vulnerability discovery in source code is allowed, but binary-based scanning and penetration tests are blocked by default, and rejected requests are automatically fallback to the older Opus 4.8.

On the other hand, he maintains restrictions in the field of biology and shows a stance not to hinder scientific research. These “domain-specific safety standards” can sometimes be a hindrance for professionals engaged in legitimate security operations, but as a remedy, a petition-based lifting framework called the Cyber Verification Program (CVP) has also been provided. In today’s world, where AI evolution is directly linked to national security and export controls, prompt guides have transformed beyond mere technical manuals into strategic documents demonstrating how to use “maximum intelligence within permitted limits” while confronting legal regulations and ethical boundaries.

Reference Page

• 【Introducing Claude Opus 5 – Anthropic】 https://www.anthropic.com/news/claude-opus-5

• 【Claude’s Constitution – Anthropic】 https://www.anthropic.com/constitution

• 【Claude Opus 5 System Card – Anthropic】 https://www.anthropic.com/claude-opus-5-system-card

• 【Models overview – Claude Platform Docs】 https://platform.claude.com/docs/en/about-claude/models/overview

[#ClaudeOpus5 #Anthropic #プロンプトエンジニアリング #AIエージェント #生成AI #科学技術 #DX]