Claude Sonnet 4.6
by Anthropic
Near-flagship intelligence at mid-tier pricing with major advances in computer use, coding, and reasoning. Released February 17, 2026.
Overview
Claude Sonnet 4.6 represents a full upgrade across multiple dimensions—coding, computer use, long-context reasoning, agent planning, knowledge work, and design. It delivers Opus-class performance for many workflows while maintaining higher speed and lower cost.
Core Capabilities
Computer Use: Major Advancement
- Interacts with computers like humans (clicking, typing, navigating)
- Works across legacy systems and modern applications (Chrome, LibreOffice, VS Code)
- Human-level capability on practical tasks (complex spreadsheets, multi-step web forms)
- 94% on insurance benchmarks (Box testing)
- 15 percentage point improvement over Sonnet 4.5 on heavy reasoning Q&A
Coding & Software Engineering
- 80.2% on SWE-bench Verified (with prompt modification, averaged over 10 trials)
- Expert-level code quality and refactoring
- Multi-file codebase navigation
- Best-in-class performance on agentic workflows
Long-Context Reasoning
- 1M token context window (beta)
- Context compaction: automatically summarizes older context
- Strong performance on Humanity’s Last Exam (complex multidisciplinary reasoning)
- Better at finding and reasoning across long contexts
Design & Visual Output
- Notably more polished visual outputs
- Better layouts, animations, and design sensibility
- Fewer iteration rounds to production-quality results
- Preferred over Opus 4.5 in 59% of user comparisons
Adaptive Thinking Features
- Intelligent decision-making on when deeper reasoning is needed
- Extended thinking for complex reasoning tasks
- Web search and fetch tools that automatically write and execute code
- Generally available tools: code execution, memory, programmatic tool calling, tool search
Performance Benchmarks
- SWE-bench Verified: 80.2% (with prompt modification)
- Humanity’s Last Exam: Strong results on complex multidisciplinary reasoning
- ARC-AGI-2: High scores with max/high effort
- OSWorld: Substantially improved computer use capabilities
- BigLaw Bench: 90.2% (legal work)
- Insurance Benchmarks: 94% accuracy
User Feedback
Early customers report:
- Strong frontend code generation
- Excellent financial analysis capabilities
- Less prone to overengineering compared to Opus 4.5
- Better instruction following and fewer hallucinations
- More consistent multi-step task completion
- Lower rate of “laziness” in responses
Pricing
- Input tokens: $3.00 per million
- Output tokens: $15.00 per million
- Significantly lower cost than Opus 4.6 while maintaining high performance
Best Use Cases
- Support automation and customer service
- Finance operations and analysis
- Internal tooling and multi-step automation
- Full-stack web development
- Code review and security analysis
- Multi-step document generation
- Extended conversations with context management
Key Advantages
- Intelligence-to-Price Ratio: Practical choice as default for many workflows
- Computer Use: Human-level interaction with digital systems
- Context Window: 1M tokens enables processing large datasets
- Cost Efficiency: Lower compute bills, faster iteration cycles
- Reliability: More consistent multi-step task completion
- Enterprise Ready: Designed for scalable automation beyond prototypes
Market Impact
Represents a fundamental shift in AI economics—high intelligence at affordable pricing makes it practical for everyday workflows rather than premium use cases only. Organizations now treat it as reliable for multi-step automation rather than prototyping tool.
Availability
- Anthropic API: Direct access
- Azure Foundry: Integration with M365 Work IQ, Fabric IQ, web access
- Google Cloud Vertex AI
- Amazon Bedrock
- Kiro: Available with credit system
Comparison to Previous Models
- vs. Opus 4.5: Faster, more efficient, preferred 59% of the time
- vs. Claude 3.5 Sonnet: Full upgrade across all dimensions
- vs. OpenAI GPT-5.2: Different architectural approach, complementary strengths