Anthropic/claude-sonnet-4-6

Supported Functionality

Item	Specification
Input	Text, Image
Output	Text
Context	200,000 tokens (1,000,000 in beta)
Max Output	64,000 tokens
Vision	✓ Supported
Function Calling	✓ Supported (computer use, code execution, memory, MCP)

Description

Claude Sonnet 4.6 is Anthropic's mid-tier flagship released on February 17, 2026, available via the Claude API as claude-sonnet-4-6 and now the default model on claude.ai and Claude Cowork for Free and Pro users. Pricing holds at $3 / $15 per million input/output tokens — identical to Sonnet 4.5 — with up to 90% savings via prompt caching and 50% via batch processing.

Sonnet 4.6's defining breakthrough is delivering near-Opus performance at Sonnet pricing: 79.6% on SWE-Bench Verified (within 1.2 points of Opus 4.6), 72.5% on OSWorld-Verified (essentially tied with Opus 4.6's 72.7%), and a 4.3× leap on ARC-AGI-2 (13.6% → 58.3%) in a single generation. In Claude Code, 70% of users preferred Sonnet 4.6 over 4.5 and 59% preferred it over the previous flagship Opus 4.5; it actually surpasses Opus 4.6 on real-world office tasks like GDPval-AA (1633 Elo) and Finance Agent (63.3%), while showing substantially improved resistance to prompt-injection attacks.

Key Capabilities

• Real-World Software Engineering: 79.6% on SWE-Bench Verified, 75.9% on SWE-Bench Multilingual; matches Opus 4.5 on long-horizon coding with fewer tokens.
• Computer Use: 72.5% on OSWorld-Verified — autonomously operates Chrome, LibreOffice, VS Code, and other GUI-only legacy software (94% accuracy on insurance vertical tasks).
• Adaptive Thinking: Four effort levels and interleaved reasoning between tool calls allow developers to tune the latency/quality trade-off precisely.
• 1M Long Context + Auto Compaction: Beta 1M-token window holds entire codebases or hundreds of contracts; Context Compaction auto-summarizes older turns to prevent hard truncation.
• Office & Financial Agents: 1633 Elo on GDPval-AA (above Opus 4.6) and 63.3% on Finance Agent — the strongest production model on real-world knowledge work today.
• Abstract Reasoning Leap: ARC-AGI-2 jumped 4.3× from 13.6% to 58.3% in one generation; ARC-AGI-1 reaches 86.5%.
• Mature Developer Ecosystem: Code execution, memory, programmatic tool calling, tool search, and code-filtered web search are all generally available, with native MCP support.

Technical Strengths

Feature	Benefit
Opus-Level Performance at Sonnet Pricing	Most coding, office, and computer-use tasks achieve near-flagship quality without paying 5× Opus pricing
1M Context + Compaction	Whole-repo analysis and very long sessions become practical; old turns compact instead of hard-truncating
Frontier Computer Use	OSWorld-Verified rose from 14.9% to 72.5% in 16 months — one of the few models reliably operating real GUIs
Adaptive + Interleaved Thinking	Model decides when and how deeply to reason inside agent loops, improving long-horizon reliability
Hardened Prompt-Injection Resistance	Major improvement over Sonnet 4.5 in computer-use and web-agent settings; on par with Opus 4.6
Anthropic Safety & Alignment	System card characterizes the model as warm, honest, prosocial with strong safety behaviors and no major misalignment concerns

Capability Ratings

Dimension	Rating	Notes
Reasoning	Excellent	ARC-AGI-2 58.3%, GPQA Diamond 74.1%; trails Opus 4.6 on the hardest scientific reasoning
Coding	Top-tier	SWE-Bench Verified 79.6% within 1.2 of Opus 4.6; long-horizon coding matches Opus 4.5
Creative Writing	Excellent	Warm, restrained, well-paced prose — Anthropic's signature style at scale
Multimodal	Strong	Text + image input; no native video/audio, optimized for UI screenshots and documents
Response Speed	Fast	Mid-tier latency, faster than Opus; thinking-off mode delivers near-instant responses
Context Window	Huge	200K default, 1M in beta; auto-compaction extends effective context further

Use Cases

• AI Coding Assistants & Claude Code: 70% of users prefer 4.6 over 4.5 — a strong default for Cursor, Claude Code, GitHub Copilot, and similar tools.
• Enterprise Desktop Automation: 72.5% OSWorld and 94% on insurance workflows make it viable for ERP, government portals, and legacy GUI software lacking APIs.
• Large Codebase & Contract Analysis: 1M context plus compaction handles entire repositories, hundreds of contracts, or dozens of research papers in a single session.
• Office & Finance Agents: Leads GDPval-AA and Finance Agent; ideal for automated report writing, financial analysis, and Excel + MCP data workflows.
• Long-Horizon Multi-Step Agents: Adaptive and interleaved thinking deliver strong results on Terminal, browser, and simulated-business tasks like Vending-Bench.
• Front-End UI & Data Reports: Customers report excellent design taste — generates production-quality pages and dashboards with minimal hand-holding.
• Enterprise-Grade Deployment: Available via Claude API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry with enterprise security and compliance controls.