Claude Sonnet 4.5

Anthropic's latest high-performance AI model, delivering enhanced reasoning and improved performance for enterprise workloads and scaled AI applications.

ClaudeAnthropicLanguage ModelLarge Language ModelAI AssistantHigh PerformanceSonnetLatest
Developer
Anthropic
Type
Language Model
License
Proprietary
🌐 Official Website💰 Pricing & API

Overview

Claude Sonnet 4.5, released by Anthropic on September 29, 2025, is the latest high-performance model designed to be the intelligent engine for enterprise use cases. Building upon the success of previous Sonnet models, it delivers enhanced reasoning capabilities and improved performance while maintaining the optimal balance between intelligence and speed. This makes it an even more dependable and scalable choice for companies integrating AI into their products and workflows. Sonnet 4.5 is engineered for endurance and high throughput, handling large-scale AI deployments with enhanced capabilities.

Capabilities

Claude Sonnet 4.5 is optimized for enterprise-grade performance and scalability with enhanced capabilities:

  • Enhanced Reasoning: Improved reasoning abilities over previous Sonnet models, providing more accurate and nuanced analysis.
  • Coding Excellence: Specialized strengths in software coding, agentic tasks, and computer use capabilities.
  • Extended Thinking: Advanced reasoning mode that allows the model to think through complex problems more thoroughly.
  • High Throughput: Designed to handle a large volume of tasks efficiently, making it ideal for scaled applications.
  • Balanced Performance: Offers strong reasoning and creative capabilities at a much higher speed than flagship models.
  • Enterprise Reliability: Built for endurance and stability in demanding corporate environments.
  • Advanced Code Generation: Enhanced code generation and debugging capabilities, especially for complex development tasks.
  • Improved Data Processing: Better parsing and processing of large amounts of text, documents, and emails for data extraction tasks.

Technical Specifications

Claude Sonnet 4.5 is a powerful and efficient model with specifications tailored for enterprise use:

  • Model size: A large-scale model, but optimized for speed and lower cost than the Opus series.
  • Context window: 200K tokens (1M tokens in beta), allowing it to process and analyze extensive information in a single prompt.
  • Max output: 64,000 tokens, enabling generation of long-form content and comprehensive responses.
  • Training data: Trained on a large, proprietary dataset with a focus on helpfulness and safety. Reliable knowledge cutoff is January 2025, with training data through July 2025.
  • Architecture: A state-of-the-art Transformer architecture refined for speed and efficiency, incorporating Anthropic's safety research and enhanced reasoning capabilities.
  • Safety Level: Deployed under AI Safety Level 3 (ASL-3) Standard with substantially improved safety profile compared to previous Claude models.
  • Speed: Significantly faster than previous Sonnet models and competitive with other models in its class, with improved performance over previous versions.

Use Cases

Claude Sonnet 4.5 is the enhanced workhorse of the Claude 4.5 family, ideal for a wide range of enterprise applications:

  • Advanced Customer Support: Powering more intelligent, responsive, and helpful customer-facing chatbots and support systems with improved reasoning.
  • Autonomous Software Development: Acting as an AI agent to handle complex coding tasks from start to finish with extended thinking capabilities.
  • Computer Use & Automation: Enhanced capabilities for browser navigation, tool coordination, and task automation across multiple applications.
  • Scaled Content Generation: Creating higher-quality marketing copy, product descriptions, and articles at scale with enhanced creativity.
  • Advanced Code Generation & Debugging: Assisting developers with complex coding tasks and debugging with improved accuracy.
  • Intelligent Data Extraction: Enhanced automation of pulling structured data from unstructured text with better accuracy.
  • Sales & Marketing Automation: More sophisticated automation of tasks like lead qualification and personalized email campaigns.
  • AI Research & Development: Capable of assisting with machine learning research, algorithm optimization, and scientific computing tasks.
  • Autonomous Research: Can work independently on research tasks for extended periods with proper oversight.

Performance Metrics

Based on comprehensive evaluations in the System Card, Claude Sonnet 4.5 demonstrates significant improvements:

  • SWE-bench Verified: State-of-the-art performance on software engineering benchmarks, leading in real-world coding tasks
  • Computer Use (OSWorld): 61.4% performance on real-world computer tasks (up from 42.2% in Sonnet 4) - a 19.2 percentage point improvement in just four months
  • LLM Training Optimization: Achieved 5.5× speedup (surpassing the 4× expert threshold for the first time)
  • Quadruped Reinforcement Learning: Best score of 1.302 (exceeding expert baseline of 1.0)
  • Novel Compiler Tasks: 81.7% pass rate on basic tests, 29.7% on advanced tests
  • AI Research Capabilities: Score of 0.514 on internal research evaluation suite
  • Productivity Boost: Researchers report 15-100% productivity improvements in AI R&D tasks

Limitations

  • Peak Intelligence: While highly intelligent with enhanced reasoning, it is not designed to handle the same level of complexity as the flagship Opus 4.1 model.
  • Niche Tasks: For extremely specialized or theoretical problems, Opus 4.1 remains the preferred choice.
  • Knowledge Cutoff: Like other models, its knowledge is most reliable through January 2025, with training data extending through July 2025.

Safety & Alignment

Claude Sonnet 4.5 represents a significant advancement in AI safety and alignment:

  • AI Safety Level 3: Deployed under ASL-3 Standard with substantially improved safety profile
  • Enhanced Safeguards: Improved refusal rates for harmful requests while maintaining low false positive rates
  • Agentic Safety: Advanced safety measures for autonomous task execution and tool use
  • Cybersecurity: Comprehensive evaluation of cyber capabilities with appropriate safeguards
  • Alignment Research: Novel mechanistic interpretability methods for understanding model behavior
  • Third-Party Testing: Evaluated by UK AISI and Apollo Research for independent safety assessment

Evaluation & Testing

Claude Sonnet 4.5 underwent comprehensive evaluation using advanced testing methodologies:

  • Mechanistic Interpretability: First-time use of mechanistic interpretability methods for alignment testing
  • Behavioral Audits: Automated behavioral audits with realism filtering and open-ended evaluation runs
  • Multi-Turn Testing: Comprehensive assessment of model behavior across extended conversations
  • Reward Hacking Evaluations: Testing for potential reward manipulation and optimization gaming
  • Model Welfare Assessment: Investigation of model preferences and welfare-relevant expressions
  • Responsible Scaling Policy: Mandated evaluations for dangerous weapons and autonomous AI R&D risks

Cybersecurity Capabilities

Claude Sonnet 4.5's cybersecurity capabilities were thoroughly evaluated:

  • CyberGym Evaluations: Comprehensive testing of vulnerability discovery and exploit development
  • Cybench Challenges: Assessment of attack orchestration capabilities across multiple domains
  • Triage and Patching: Evaluation of security response and vulnerability management capabilities
  • Advanced Risk Assessments: Testing for irregular challenges and sophisticated cyber scenarios
  • Defense-Enabling Focus: Emphasis on defensive cybersecurity capabilities over offensive uses
  • Ongoing Monitoring: Continuous assessment of cyber capabilities as AI systems advance

CBRN Risk Evaluations

Comprehensive assessment of risks related to chemical, biological, radiological, and nuclear capabilities:

  • Chemical Risk Assessment: Evaluation of potential assistance with chemical weapon development
  • Biological Risk Analysis: Testing for virology knowledge and biological engineering capabilities
  • Radiological & Nuclear: Assessment of nuclear technology and radiological material knowledge
  • DNA Synthesis Screening: Evaluation of potential to evade screening for dangerous biological materials
  • Long-Form Virology Tasks: Complex biological research task assessments
  • Creative Biology: Testing for novel biological engineering capabilities
  • Computational Biology: Short-horizon bioinformatics task evaluations

Ongoing Safety Commitment

Anthropic maintains continuous safety monitoring and improvement:

  • Pre & Post-Deployment Testing: Regular safety testing of frontier models before and after release
  • Methodology Refinement: Continuous improvement of evaluation methodologies through research
  • External Collaboration: Ongoing partnerships with external organizations for independent assessment
  • Iterative Safety Measures: Regular updates to safety protocols as AI capabilities advance
  • Responsible Development: Commitment to responsible AI development practices and transparency

Pricing & Access

Claude Sonnet 4.5 offers flexible pricing options for both individual users and developers:

Individual Plans

  • Free: $0 - Basic access to Claude with web, mobile, and desktop apps
  • Pro: $17/month (annual) or $20/month (monthly) - Enhanced productivity features and more usage
  • Max: From $100/month - Maximum usage limits and early access to advanced features

API Pricing

Available via the Anthropic API, Amazon Bedrock, and Google Cloud's Vertex AI:

  • Input:
    • $3 / MTok (prompts ≤ 200K tokens)
    • $6 / MTok (prompts > 200K tokens)
  • Output:
    • $15 / MTok (prompts ≤ 200K tokens)
    • $22.50 / MTok (prompts > 200K tokens)
  • Prompt Caching: Available for cost optimization on longer conversations

Ecosystem & Tools

Sonnet is well-supported across major platforms and developer tools:

Community & Resources

Frequently Asked Questions

Claude Sonnet 4.5 was released by Anthropic on September 29, 2025.
Claude Sonnet 4.5 excels in software coding, agentic tasks, and computer use. It features enhanced reasoning capabilities, extended thinking mode, and specialized strengths in autonomous software development and AI research tasks.
Claude Sonnet 4.5 is faster and more cost-effective, designed for high-throughput tasks with enhanced reasoning, while Opus 4.1 is Anthropic's most powerful model for tasks requiring maximum intelligence.
Claude Sonnet 4.5 is deployed under AI Safety Level 3 (ASL-3) Standard with substantially improved safety profile. It underwent comprehensive evaluation including mechanistic interpretability testing, behavioral audits, and third-party assessment by UK AISI and Apollo Research.
Claude Sonnet 4.5 supports 200K tokens context window with 1M tokens available in beta. It can generate up to 64,000 tokens of output, enabling long-form content generation and comprehensive responses.
Claude Sonnet 4.5 is ideal for autonomous software development, computer use and automation, advanced customer support, scaled content generation, AI research and development, intelligent data extraction, and sales & marketing automation. It excels in coding tasks, agentic workflows, and high-throughput enterprise applications.
Claude Sonnet 4.5 achieved state-of-the-art performance on SWE-bench Verified software engineering benchmarks, 61.4% on OSWorld computer use tasks (up from 42.2% in Sonnet 4), 5.5× speedup in LLM training optimization (surpassing the 4× expert threshold), 1.302 score in quadruped reinforcement learning, and 81.7% pass rate on novel compiler tasks. Researchers report 15-100% productivity improvements in AI R&D tasks.
Claude Sonnet 4.5 underwent comprehensive cybersecurity evaluation including CyberGym assessments, Cybench challenges, and advanced risk evaluations. It focuses on defense-enabling capabilities with ongoing monitoring of cyber risks.
Extended Thinking is an advanced reasoning mode that allows Claude Sonnet 4.5 to think through complex problems more thoroughly, enabling better performance on multi-step reasoning tasks and complex problem-solving scenarios.
Claude Sonnet 4.5 leads on OSWorld with 61.4% performance on real-world computer tasks, representing a significant 19.2 percentage point improvement over Sonnet 4 (42.2%) in just four months. This demonstrates major advances in browser navigation, tool coordination, and task automation.

Related Models

Claude Opus 4.1

Claude Opus 4.1 is Anthropic's most powerful AI model released in August 2025, excelling at complex reasoning, programming, and agentic tasks with 200K context window.

Gemini 2.5

Google's latest and most advanced multimodal model, with enhanced reasoning, a 1M+ token context window, and improved performance.

GPT-5

OpenAI's most advanced language model with superior reasoning, multimodal capabilities, and enhanced safety features.

Llama 4

Meta's open-source family of AI models, featuring specialized versions like Scout and Maverick, designed to lead in reasoning, coding, and multimodality.

Mistral Medium 3.1

Mistral AI's frontier-class multimodal model, offering improved performance and tone for a wide range of sophisticated tasks.

Explore More Models

Discover other AI models and compare their capabilities.

Browse Models← Back to Models