Claude Opus 5 - Anthropic's latest flagship model

What is Claude Opus 5

Claude Opus 5 is Anthropic's latest flagship model, positioned as "approaching Fable 5 's frontier intelligence at about half the price", and compared to Opus 4.8 it greatly improves performance at unchanged cost, and has become the default model for Claude Max and the strongest choice on Claude Pro. It introduces adjustable effort thinking levels, letting users trade off between stronger reasoning and lower latency and fewer tokens; it tops multiple benchmarks in programming, knowledge work, business automation and computer operation such as Frontier-Bench, GDPval-AA, ARC-AGI 3, AutomationBench, and OSWorld 2.0, trailing Mythos 5 only in a few tasks such as cybersecurity.

Main features of Claude Opus 5

• Adjustable effort thinking levels: Supports levels such as low, medium, high, xhigh, and max; higher levels yield stronger reasoning, while lower levels are faster and save more tokens, making it easy to balance quality, speed, and cost per task.
• Agentic programming capability: Aimed at terminal and software engineering tasks, it scores 43.3% on Frontier-Bench v0.1, higher than Fable 5, Opus 4.8, and GPT-5.6 Sol; officially stated to more than double performance compared to Opus 4.8 with lower per-task cost.
• Code collaboration optimization: At the highest level of CursorBench 3.2, it is only 0.5% behind Fable 5's peak, at about half the cost; at high, xhigh, and max levels it leads other models at the same budget.
• Knowledge work and writing: Covers analysis, research, documents, and office tasks; scores 1861 on GDPval-AA v2, leading Fable 5, GPT-5.6 Sol, and Opus 4.8.
• Smart search and browsing: Has agentic search capability, scoring 90.8% on BrowseComp, slightly above GPT-5.6 Sol's 90.4%, and can be used for complex information retrieval and synthesis.
• Computer operation capability: Scores 70.6% on OSWorld 2.0, higher than Fable 5 and GPT-5.6 Sol; officially stated to surpass Fable 5's best result at about one-third of the cost.
• Business process automation: Can complete business tasks end to end, with a 26.0% pass rate on AutomationBench, about 1.5 times that of the next-best model at the same cost; even at the lowest effort level it completes more tasks than all competitors.
• Novel problem solving: Scores 30.2% on ARC-AGI 3, about three times that of the runner-up GPT-5.6 Sol, emphasizing the ability to handle tasks not seen during training.
• Scientific research assistance: Comprehensively surpasses Opus 4.8 in life science evaluations, covering structural biology, organic chemistry, and bioinformatics; improves spectral inference of molecular structures by 10.2 percentage points and prediction of functional impact of protein sequence variants by 7.7 percentage points.
• More proactive and also more prudent: Officially positioned as a more proactive and more thoughtful model that can reduce back-and-forth confirmation, self-check, and recover from errors during tasks; it also emphasizes safety guardrails.
• High-frequency daily use and product integration: Runs more efficiently and is suitable for daily use; it has become the default model for Claude Max and is also the highest-capability option available on Claude Pro, with interface entries for levels such as "Opus 5 High".

How to use Claude Opus 5

• Access the platform: Open the Claude app or web version and switch to Claude Opus 5 in the model selector.
• Use according to subscription tier: Claude Max already uses Opus 5 as the default model; Claude Pro users can select it as the current highest-capability option.
• Adjust the effort level by task: For simple Q&A, summaries, and polishing, use a low level for faster and cheaper results; for code debugging, complex analysis, and research tasks, switch to high, xhigh, or max for stronger reasoning.
• Give clear goals and constraints: Specify the deliverables, format, word count, deadline, available tools, and acceptance criteria to reduce back-and-forth and leverage its "more proactive" nature.
• Programming scenarios: Paste the repository structure, error logs, API documentation, and expected behavior. Opus 5's strongest areas are locating problems, modifying code, running tests, and explaining trade-offs.
• Knowledge work: Provide background material, target audience, and output templates, then conduct research, outlining, first drafts, proofreading, and rewriting; GDPval-AA v2 shows it leads in knowledge work.
• Search and synthesis: Require "retrieve first, then summarize, cite sources, and distinguish facts from speculation," corresponding to its agentic search capability; BrowseComp score of 90.8%.
• Automated workflows: Break repetitive office workflows into steps, inputs, outputs, and exception handling, and let it execute end to end; AutomationBench shows its same-cost pass rate is about 1.5 times that of the runner-up.
• Computer operation tasks: Clarify which software is allowed to be operated, the file scope, and prohibited actions, then let it execute; Opus 5 scores 70.6% on OSWorld 2.0.
• Scientific research scenarios: Use it for structural biology, organic chemistry, bioinformatics, and other problems, such as inferring molecular structures from spectra and analyzing the functional impact of protein variants; however, expert review is still required.
• Control costs: First trial-run with a low level, then upgrade for difficult tasks; split long tasks into smaller batches and prioritize reserving max effort for key hard problems, because higher levels are smarter but consume more tokens.

Claude Opus 5's core advantages

• Half price approaching flagship: Intelligence level is close to Fable 5, but the price is about half of the latter; CursorBench 3.2's highest level differs from Fable 5's peak by only 0.5%, yet the cost is about half.
• More for the same price: Compared with Opus 4.8, it greatly improves performance at unchanged cost; the official claim is that Frontier-Bench programming task performance more than doubles while the per-task cost is lower.
• Most outstanding programming capability: Frontier-Bench v0.1 score of 43.3%, higher than Fable 5's 33.7%, GPT-5.6 Sol's 34.4%, and Opus 4.8's 21.1%.
• More competitive at the same budget: At the high, xhigh, and max levels, it leads other models in score at the same cost; the official narrative can be summarized as "spend half the money to complete nearly 99% of the work."
• Scarce effort flexibility: One model covers multiple levels of need from fast and economical to deep reasoning; low levels can be used frequently for daily tasks, high levels for tackling hard problems, and the cost curve is more controllable.
• Leading in knowledge work: GDPval-AA v2 score of 1861, surpassing Fable 5, GPT-5.6 Sol, and Opus 4.8; suitable for research, analysis, writing, and office decision-making.
• Leap in solving new problems: ARC-AGI 3 score of 30.2%, about three times that of the second-place GPT-5.6 Sol, indicating stronger generalization ability when facing unfamiliar tasks.
• High cost-effectiveness in automation and computer operation: AutomationBench's pass rate at the same cost is about 1.5 times that of the runner-up; OSWorld 2.0 surpasses Fable 5's best result at about one-third of the cost.
• Search and synthesis capabilities rank near the top: BrowseComp 90.8%, slightly higher than GPT-5.6 Sol's 90.4%, suitable for complex retrieval, document comparison, and review generation.
• Clear gains in scientific research capabilities: Surpasses Opus 4.8 across all life science evaluations, with the largest improvement in organic chemistry; spectral inference of molecular structures +10.2 percentage points, prediction of the functional impact of protein variants +7.7 percentage points.

Claude Opus 5's project address

• Claude Opus 5：https://www.anthropic.com/news/claude-opus-5

Comparison of Claude Opus 5 with similar competing products

Claude Opus 5's application scenarios

• Software development and refactoring: Suitable for requirements clarification, code generation, bug localization, test completion, refactoring explanations, and terminal command assistance; Frontier-Bench terminal programming ranks first at 43.3%, making it its strongest scenario.
• Routinization of AI programming assistants: In Cursor-like workflows, it handles long-task decomposition, cross-file modifications, and regression checks; the highest tier of CursorBench approaches Fable 5's peak but at about half the cost, making it suitable for high-frequency calls.
• Enterprise knowledge work: Writing proposals, doing research, organizing meeting minutes, generating reports, editing resumes/official documents/business documents; leads GDPval-AA v2 at 1861, suitable for high-frequency office use.
• Deep retrieval and review: Given a topic, automatically retrieves, compares sources, summarizes conclusions, and flags uncertainty; BrowseComp 90.8%, suitable for competitor research, industry research, and literature review.
• Business process automation: Runs processes such as form processing, data transfer, approval drafts, customer follow-up, and report generation end to end; AutomationBench pass rate at the same cost is about 1.5 times that of the next best.
• Computer operation and lightweight RPA: Within clearly defined permission boundaries, operates software, organizes files, and handles spreadsheets and web processes in batches; OSWorld 2.0 70.6%, with better cost efficiency than Fable 5.

Frequently asked questions about Claude Opus 5