Claude Opus 5

Claude Opus 5: Model Introduction & Practical Guide

Claude Opus 5 is Anthropic's flagship for everyone - described by the company as "frontier intelligence near Fable 5 at about half the price." It is now the default model on Claude Max, the top capability on Claude Pro, and the model that redefined what the Opus tier means: near-frontier capability with an adjustable reasoning dial instead of a fixed price/quality point.

Here's the short version: Opus 5 closes most of the gap to the Mythos-class Fable tiers while costing roughly half, and it tops multiple independent-adjacent benchmarks in the process: GDPval-AA v2 at 1861 (leading knowledge work), BrowseComp at 90.8% (agentic search), OSWorld 2.0 at 70.6% (computer use), ARC-AGI 3 at 30.2% (novel problems, ~3x the runner-up), Frontier-Bench v0.1 at 43.3% (agentic coding), and AutomationBench at 26.0% (business automation). The five-level effort control (low → max) lets teams trade cost for depth per task. The trade-offs: Fable 5 still leads on the absolute longest-horizon reliability and a handful of security-domain benchmarks, the new tokenizer's output-length changes can shift costs on legacy prompts, and max-effort runs consume tokens quickly.

This guide covers model overview, core features, technical specifications, capability comparison, core advantages, recommended use cases, example prompts, and selection recommendations.

Quick Facts

Table of Contents

1. Model Overview
2. Core Features
3. Technical Specifications
4. Capability Comparison
5. Core Advantages
6. Recommended Use Cases
7. Example Prompts
8. Selection Recommendations
9. FAQ
10. Sources & Further Reading

1. Model Overview

Anthropic's Opus 5 launch framed the model in explicitly economic terms: near-Fable-5 intelligence at roughly half the price, with better performance than Opus 4.8 at the same cost. That combination - capability step-up without a price step-up - is why it immediately became Claude Max's default and Claude Pro's strongest option.

The capability evidence spans five domains. In agentic coding, Frontier-Bench v0.1 lands at 43.3% - above Fable 5, Opus 4.8, and GPT-5.6 Sol - with Anthropic claiming over double Opus 4.8's performance at lower per-task cost. In knowledge work, GDPval-AA v2 scores 1861, leading Fable 5, GPT-5.6 Sol, and the previous Opus. In agentic search, BrowseComp 90.8% edges GPT-5.6 Sol's 90.4%. In computer use, OSWorld 2.0 reaches 70.6%, ahead of Fable 5 and GPT-5.6 Sol at roughly a third of the cost of Fable 5's best. And in novel-problem reasoning, ARC-AGI 3 hits 30.2% - about three times the runner-up - which speaks to handling tasks outside training distributions.

The effort dial is the operational centerpiece. Five levels (low, medium, high, xhigh, max) span from fast/cheap to maximum reasoning, and Anthropic notes that even the lowest effort level completes more tasks than competitors on some suites. For scientific work, Opus 5 improves across life sciences (structural biology, organic chemistry, bioinformatics), with specific gains in spectral-to-structure inference (+10.2 points) and protein variant effect prediction (+7.7 points).

Anthropic also emphasizes disposition: Opus 5 is positioned as more proactive and deliberate - fewer confirmation loops, more self-checking, better recovery from its own mistakes - while shipping with strengthened safety guardrails.

2. Core Features

Adjustable effort. Five levels from low to max - simple tasks run fast and cheap, hard tasks get maximum reasoning depth.

Agentic coding. Frontend-, terminal-, and repository-level work at Frontier-Bench-leading performance (43.3%), with CursorBench 3.2 performance within 0.5% of Fable 5's peak at about half the cost.

Knowledge work and writing. Analysis, research, documents, and office tasks at GDPval-AA v2 1861 - the leading score.

Agentic search and browsing. BrowseComp 90.8% for complex information retrieval and synthesis.

Computer use. OSWorld 2.0 at 70.6% - graphical interface operation for cross-application tasks.

Business process automation. AutomationBench 26.0% pass rate at roughly 1.5x the next-best model per unit cost.

Scientific research. Life-science gains across structural biology, organic chemistry, and bioinformatics.

Proactive, careful disposition. Fewer confirmation cycles, stronger self-checking, and recovery from mid-task errors.

3. Technical Specifications

4. Capability Comparison

Positioning read. Opus 5's pitch is unusual: it beats its more expensive sibling on several headline benchmarks (GDPval, BrowseComp, OSWorld) while costing roughly half. Fable 5's remaining advantages concentrate in the longest-horizon reliability and security-domain capability depth. For most teams, Opus 5 is now the default frontier choice, with Fable 5 reserved for the work that justifies it.

5. Core Advantages

1. Frontier-adjacent at half price. The capability-per-dollar step that reshapes model routing.
2. Leaderboard breadth. Top scores across coding, knowledge work, search, and computer use simultaneously.
3. The effort dial. Five levels for per-task cost/quality tuning - even low-effort runs complete broadly.
4. Proactive disposition. Fewer confirmation loops and better self-recovery reduce supervision overhead.
5. Scientific depth. Measurable life-science gains for research workflows.
6. Ecosystem defaults. Standard on Claude Max, top choice on Pro, ready in Claude Code.

6. Recommended Use Cases

• Agentic software development: repository-scale coding, terminal automation, code review.
• Knowledge-work deliverables: research, analysis, reports, and documents.
• Computer-use automation: GUI operation across applications.
• Business process automation: multi-step enterprise workflows.
• Scientific research assistance: structural biology, chemistry, and bioinformatics tasks.
• Cost-controlled frontier work: applying max effort selectively while defaulting to lower levels.

7. Example Prompts

1. Agentic coding with effort control

2. Knowledge-work deliverable

3. Computer-use task

4. Automation flow

5. Research assistance

8. Selection Recommendations

Choose Claude Opus 5 if:

• You want near-frontier capability at roughly half the flagship premium.
• Your mix spans coding, knowledge work, search, and computer use.
• You want per-task effort control to manage cost.
• You value proactive, low-supervision agent behavior.
• You are on Claude Max or Pro and want the strongest available option.

Upgrade to Fable 5 if:

• The work demands the absolute ceiling on long-horizon reliability.
• Security-domain depth (cyber/bio) is essential.

Consider other families if:

• You need open weights (DeepSeek, GLM, Kimi K3).
• You need the largest context windows (Gemini/GPT 1.5M-class).

Sources & Further Reading

• Claude Opus 5 - AI Toolset (Chinese overview)
• Claude models - Anthropic

Benchmark figures are as published by Anthropic and third-party summaries at review time. Effort-level performance and pricing vary by task; validate against your workloads before routing production traffic.