Claude Opus 5 – Anthropic's Latest Flagship Model

1. What is Claude Opus 5

Claude Opus 5 is a flagship large language model released by Anthropic in 2025, positioned as a high-value option with "near-frontier intelligence at about half the cost." Under the same cost structure as Opus 4.8, Opus 5 delivers a significant performance leap, becoming the default model for Claude Max subscriptions and the highest capability option available on Claude Pro. By introducing adjustable effort thinking levels, Opus 5 enables users to flexibly balance between stronger reasoning, lower latency, and reduced token usage, covering a wide range of scenarios from high-frequency daily tasks to in-depth professional analysis. Opus 5 has achieved leading results in multiple benchmarks, including Frontier-Bench, GDPval-AA, ARC-AGI 3, AutomationBench, and OSWorld 2.0, with particularly strong performance in programming, knowledge work, business automation, and computer operations.

Image source: Official article
Image source: official article

Technical Positioning and Domain: Belongs to the fields of natural language processing and general artificial intelligence, positioned as a high-value flagship large language model. It is aimed at individual developers, enterprise teams, and high-frequency daily users who require near-frontier intelligence but have limited budgets. Its unique feature lies in the effort level mechanism, which enables the same model to flexibly cover both lightweight and fast responses as well as deep reasoning.

Development Background: Developed by the Anthropic team, building upon the accumulation from previous versions such as Opus 4.8. Anthropic has deep technical expertise in balancing safety, interpretability, and model capabilities. Opus 5 is a product of their "more proactive and thoughtful" design philosophy, emphasizing reduced back-and-forth confirmation, self-checking, and error recovery capabilities.

Core Value: Resolves the long-standing contradiction between "high cost of frontier models" and "limited capabilities of low-cost models." Opus 5 achieves near-peak capabilities at about half the price of Fable 5, while outperforming Opus 4.8 and even surpassing some metrics of Fable 5 in multiple benchmark tests. This allows more users to access flagship-level intelligence at a reasonable cost.

2. Key Features

• Adjustable Effort Thinking Modes: Supports five levels of thinking intensity: low, medium, high, xhigh, and max. Higher levels provide deeper reasoning and greater accuracy, while lower levels significantly reduce latency and token consumption. This flexible design enables the same model to handle simple Q&A tasks as well as complex programming and research tasks, greatly optimizing cost control.

• Agentic Programming Capabilities: Designed for end-user and software engineering tasks, achieving a score of 43.3% on Frontier-Bench v0.1, surpassing Fable 5 (33.7%) and GPT-5.6 Sol (34.4%). According to the official statement, it outperforms Opus 4.8 by more than double, with lower single-task costs, making it ideal for scenarios such as requirement clarification, code generation, bug localization, and test completion.

• Code Collaboration Optimization: At the highest level of CursorBench 3.2, Opus 5 is only 0.5% behind Fable 5, but at about half the cost. In high, xhigh, and max modes, Opus 5 achieves higher scores than other models under the same budget, making it an ideal choice for daily use as an AI programming assistant.

• Knowledge Work and Writing: Covers tasks such as analysis, research, documentation, and office work. It scores 1861 on GDPval-AA v2, outperforming Fable 5 (1747) and GPT-5.6 Sol (1736). Suitable for high-frequency knowledge-intensive tasks like proposal writing, meeting minutes organization, and business document generation.

• Intelligent Search and Browsing: Features agentic search capabilities, scoring 90.8% in the BrowseComp evaluation, slightly higher than GPT-5.6 Sol's 90.4%. Ideal for complex document retrieval, cross-source information comparison, and comprehensive summary generation, it can automatically search, cite sources, and distinguish between facts and speculation.

• Computer Operation Capabilities: Scores 70.6% on OSWorld 2.0, outperforming Fable 5 (66.1%) and GPT-5.6 Sol (62.6%). The official claims it can achieve results comparable to the best performance of Fable 5 at about one-third of the cost, suitable for lightweight RPA scenarios such as operating software, organizing files, and batch processing spreadsheets within defined permission boundaries.

• Business Process Automation: Can complete business tasks end-to-end, with a pass rate of 26.0% on AutomationBench, approximately 1.5 times that of the second-best model at the same cost. Even in the lowest effort mode, it outperforms all competitors in the number of tasks completed, making it suitable for repetitive office workflows such as form processing, data migration, and approval drafts.

• Novel Problem Solving: Scores 30.2% on ARC-AGI 3, about three times that of the second-place GPT-5.6 Sol (7.8%). Emphasizes the ability to handle tasks not seen during training, demonstrating strong generalization capabilities on unfamiliar problems.

• Scientific Research Assistance: Outperforms Opus 4.8 comprehensively in life sciences evaluations, covering structural biology, organic chemistry, and bioinformatics. It improves molecular structure inference from spectra by 10.2 percentage points and enhances prediction of the impact of protein sequence variations by 7.7 percentage points. However, expert verification is still required.

3. How to Use

1. Access the Platform: Open the Claude app (web or desktop) and switch to Claude Opus 5 in the model selector. Claude Max subscribers have this as their default model; Claude Pro users can select it as the current highest capability. You will see an entry similar to "Opus 5 High" in the interface.

2. Adjust the Effort Level Based on the Task: For simple Q&A, summarization, polishing, and similar tasks, use the low or medium effort level to achieve faster response speeds and lower token consumption. For code debugging, complex analysis, and research tasks, switch to high, xhigh, or max effort levels to gain stronger reasoning capabilities. It is recommended to first run the task on a lower effort level to confirm the direction before increasing the effort level.

3. Provide Clear Goals and Constraints: Clearly specify the deliverables, format, word count, deadline, available tools, and acceptance criteria in your prompt. Opus 5 is designed to be a "more proactive" model, reducing the need for back-and-forth confirmation, so clear goals can fully leverage its self-planning and error recovery capabilities.

4. Best Practices for Programming Scenarios: Paste the repository structure, error logs, API documentation, and expected behavior. Opus 5 excels in identifying issues, modifying code, running tests, and explaining technical trade-offs. It can be combined with tools like Cursor to break down long tasks into smaller batches for execution.

5. Knowledge Work and Search: Provide background materials, target audience, and output templates, and specify "search first, then summarize, cite sources, and distinguish facts from speculation." This aligns with its agentic search capabilities, and its 90.8% score on BrowseComp indicates its suitability for complex research and summary generation.

6. Automation and Computer Operations: Break down repetitive office workflows into steps, inputs, outputs, and error handling. Clearly define the allowed software, file scope, and prohibited actions. Opus 5's performance on AutomationBench and OSWorld shows its ability to execute workflows end-to-end, but be mindful of security boundaries.

7. Cost Control: First run tasks on a low effort level, then increase the effort level for challenging tasks. Break down long tasks into smaller batches for execution, and prioritize using the max effort level for critical challenges. Higher effort levels are more intelligent but consume more tokens, so allocate them wisely based on your budget.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: For users who seek "near-flagship capabilities + cost control" in programming, knowledge work, automation, and daily research tasks, Claude Opus 5 is currently the most cost-effective option. Its effort elasticity mechanism allows the same model to adapt to a variety of tasks ranging from simple Q&A to deep reasoning, and it leads in several benchmarks under the same cost. If budget is sufficient and extreme peak performance is required (e.g., cybersecurity, some complex reasoning tasks), Claude Fable 5 or Mythos 5 may be more suitable. For teams already deeply integrated with the OpenAI ecosystem (e.g., deploying large-scale systems using the GPT API), GPT-5.6 Sol still has advantages in DeepSWE and certain health evaluations, but its overall cost-effectiveness is lower than that of Opus 5. Users are advised to prioritize benchmark scores relevant to their high-frequency scenarios and combine them with practical testing for evaluation.

6. Editor's Summary

Claude Opus 5 represents an important strategic shift in Anthropic's flagship model positioning: rather than solely pursuing dominance in peak capabilities, it leverages an effort elasticity mechanism and cost control strategies to bring cutting-edge intelligence to a broader user base at a lower entry threshold. From a technical perspective, its core innovation lies in the design of "adjustable thinking depth"—a concept not entirely new in the industry (such as variations of chain-of-thought)—but Opus 5 systematizes it into a five-tier framework ranging from low to max, while maintaining competitiveness across all tiers. This requires the model to possess strong generalization and robustness capabilities at both the training and inference levels. Performance scores of 43.3% on Frontier-Bench and 30.2% on ARC-AGI 3 also confirm its solid foundation in programming and problem-solving for new tasks.

In terms of practical value, Opus 5's most notable contribution is breaking the industry's conventional belief that "high performance necessarily implies high cost." Achieving near-peak capabilities at about half the price of Fable 5, while leading in multiple benchmarks under similar budgets, makes it accessible for mid-sized teams and individual developers to obtain flagship-level intelligent assistance. Its comprehensive capabilities in knowledge work, search, and automation also make it a "versatile" daily tool, rather than a specialized model limited to specific domains.

In terms of target users, Opus 5 is most suitable for the following three categories: first, software engineers who frequently use AI programming assistants (especially in IDE integration scenarios like Cursor); second, corporate professionals and researchers engaged in substantial knowledge work (analysis, research, writing, office decision-making); and third, operations teams aiming to achieve commercial process automation and lightweight RPA with controllable costs. For users seeking extreme peak performance with ample budgets, Fable 5 or Mythos 5 may still be necessary complements.

Looking ahead, Opus 5's effort mechanism provides a blueprint for dynamically adjusting model capabilities and costs, a feature that could be adopted by more models in the future. Anthropic's ongoing investment in safety guardrails and "more proactive, thoughtful" design also gives it an edge in compliance and reliability. As the Claude ecosystem expands (e.g., API openness and third-party integrations), the cost-performance advantage of Opus 5 is likely to grow further. Overall, this is a product that demonstrates a highly mature technical strategy and commercial positioning, making it a key focus for future attention.

7. Application Scenarios

• Software Development and Refactoring: Ideal for requirement clarification, code generation, bug localization, test completion, refactoring explanation, and terminal command assistance. Opus 5 leads in the Frontier-Bench terminal programming task with a 43.3% success rate, making it its strongest scenario. Developers can paste repository structures, error logs, and API documentation, allowing Opus 5 to identify issues, modify code, and explain technical trade-offs.

• Daily Use of AI Programming Assistants: Perform long-task decomposition, cross-file modifications, and regression checks within Cursor-like workflows. While the top tier of CursorBench is close to Fable 5's peak, Opus 5 achieves this at about half the cost, making it suitable for high-frequency use. Developers can break down large refactoring tasks into multiple small steps and have Opus 5 execute them incrementally.

• Enterprise Knowledge Work: Useful for drafting proposals, conducting research, organizing meeting minutes, generating reports, and revising resumes, official documents, and business reports. GDPval-AA v2 leads with a score of 1861, making it well-suited for frequent use in office environments. Users only need to provide background information, target audience, and output templates, and Opus 5 can complete the research, outline, draft, and proofreading.

• Deep Retrieval and Literature Review: After being given a topic, the model automatically retrieves sources, compares them, summarizes conclusions, and annotates uncertainties. With a 90.8% success rate on BrowseComp, it is ideal for competitive analysis, industry research, and literature organization. Users can instruct the model to "first retrieve, then summarize, cite sources, and distinguish between facts and speculation," fully leveraging its agentic search capabilities.

• Business Process Automation: Enable end-to-end execution of processes such as form processing, data migration, draft approvals, customer follow-ups, and report generation. At the same cost, Opus 5's pass rate on AutomationBench is approximately 1.5 times that of the second-best model. Users must break down the process into steps, inputs, outputs, and error handling, allowing Opus 5 to execute it end-to-end.

• Computer Operation and Lightweight RPA: Operate software, organize files, batch process spreadsheets, and automate web workflows within clearly defined permission boundaries. Opus 5 scores 70.6% on OSWorld 2.0 and offers better cost efficiency than Fable 5. Users must specify the allowed software, file scope, and prohibited actions before the model can execute the tasks.

• Scientific Research Assistance: Applied to problems in structural biology, organic chemistry, and bioinformatics, such as inferring molecular structures from spectra or analyzing the functional impact of protein mutations. Opus 5 outperforms Opus 4.8 comprehensively in life sciences evaluations, but expert verification of results is still required.

8. FAQ

Q: How to choose between Claude Fable 5 and Opus 5?
A: Choose Fable 5 if you require extreme peak performance and have a sufficient budget. Choose Opus 5 if you prioritize cost-effectiveness and need high performance for frequent daily tasks. Under the highest tier of CursorBench, Opus 5 is only 0.5% behind Fable 5 in peak performance, but at about half the cost. For most scenarios such as programming, knowledge work, and automation, Opus 5 is more than sufficient.

Q: What improvements does Opus 5 have over Opus 4.8?
A: According to official statements, Opus 5 maintains the same cost while significantly improving performance. In the Frontier-Bench programming tasks, the score increased from 21.1% to 43.3%, representing more than double the performance improvement with lower cost per task. There are also notable improvements in knowledge work, research, and search dimensions.

Q: How should I understand the current model hierarchy of Claude?
A: Roughly speaking, Opus 5 is the flagship model offering high cost-effectiveness, Fable 5 is a more advanced tier, and Mythos 5 excels in a few highly complex areas (such as cybersecurity). Opus 5 is the default model for Claude Max and the highest capability option available for Claude Pro users.

Q: Where can I use Claude Opus 5?
A: Claude Max has already set Opus 5 as the default model; Claude Pro users can select it as the highest capability option. You can find an entry similar to "Opus 5 High" in the interface. Currently, it is only available through the Claude app and does not support local deployment or third-party APIs.

Q: What is the "effort" tier, and how do I use it?
A: The "effort" tier refers to the intensity of the model's thinking process. Increasing the effort level makes the model "smarter" (with deeper reasoning), while decreasing it makes it faster and more token-efficient. Use low/medium for simple tasks, and high/xhigh/max for complex tasks such as coding, research, and automation. It is recommended to first test with lower tiers to determine the direction, then increase the tier for fine-tuning.

Q: What is it best at?
A: It excels most in software engineering and terminal programming (43.3% on Frontier-Bench), followed by knowledge work (GDPval-AA 1861), intelligent search (BrowseComp 90.8%), business process automation (AutomationBench 26.0%), computer operation (OSWorld 70.6%), and solving new problems (ARC-AGI 30.2%). It has achieved top performance in multiple benchmark tests.

Q: Is it reliable in research scenarios?
A: In life sciences evaluations, it outperforms Opus 4.8 comprehensively. It improves molecular structure inference from spectra by 10.2 percentage points and predicts the functional impact of protein mutations with an improvement of 7.7 percentage points. However, the model may still produce incorrect inferences, and formal research work must be reviewed by domain experts.

Q: How to control usage costs?
A: Start with lower tiers to test, and only increase the tier for difficult tasks. Break long tasks into smaller batches, and reserve max effort for critical challenges. Higher effort tiers are smarter but consume more tokens, so it's important to allocate them reasonably based on your budget. Claude Max subscriptions offer a fixed quota, while Claude Pro uses a pay-as-you-go model.

9. Project Links

• Official Release Page: https://www.anthropic.com/news/claude-opus-5
• Claude Product Website: https://claude.ai (Access Opus 5 through this platform; requires a Claude Max or Claude Pro subscription)