Claude Opus 4

Claude Opus 4 is Anthropic’s hybrid-reasoning frontier model that inaugurated the Claude 4 model generation. Built to handle long-running, thousands-step engineering tasks, Opus 4 introduced native extended thinking alongside a near-instant response mode, setting a benchmark for sustained, multi-hour autonomous agent execution.

Model Overview

Opus 4 was engineered specifically for heavy-cognition software engineering, long-horizon tool execution, and complex system refactoring. It operates as both a near-instant conversational engine and a deep-reasoning agent capable of managing multi-file edits without losing structural coherence.

• Developer: Anthropic
• Model Tier: Opus (Frontier / Maximum Intelligence)
• Release Date: May 22, 2025
• Architecture: Hybrid Transformer with Dual Execution Modes (Standard & Extended Thinking)
• Primary Target Use Cases: Autonomous codebase refactoring, multi-file bug tracing, sustained agentic execution, and scientific/mathematical reasoning.

Core Capabilities & Performance Profile

Extended Thinking & Dual Reasoning

Opus 4 operates across two distinct execution pathways: 1. Standard Mode: High-speed generation for low-latency chat and direct tool calls. 2. Extended Thinking Mode: Allocates an internal, hidden reasoning budget (up to 64k tokens) to analyze complex logic, evaluate edge cases, and run multi-path planning before producing final output.

Sustained Agentic Coding

• Multi-Hour Continuity: Capable of working continuously over thousands of agentic loop steps to resolve complex repository issues.
• SWE-Bench Leadership: Established a 72.5% score on SWE-Bench Verified at release, outperforming previous-generation models on real-world open-source pull request resolution.
• Minimal Scaffold Execution: Operates using simple tool scaffolds (bash and string-replacement file tools), relying on internal model reasoning rather than heavy external agent framework logic.

Persistent Local Memory Files

When provided with local file system access, Opus 4 generates, updates, and reads dedicated “memory files” to maintain state, context, and project goals across long-running developer sessions.

Technical Specifications