claude-sonnet-4-5-20250929

Basic Information

Claude Sonnet 4.5 is the latest-generation AI model officially released by Anthropic on September 29, 2025, built on Anthropic’s self-developed technical architecture. It represents a significant upgrade to the Claude series, with the model API ID claudesonnet-4-5-20250929, marking a major technological breakthrough by Anthropic in the fields of AI programming and complex agent development. Ordinary users can access it via the web interface, as well as through the Claude.ai app on iOS and Android. Developers can invoke it through the Claude Developer Platform, Amazon Bedrock, and Google Cloud’s Vertex AI. The knowledge cutoff date is January 2025, and the training data cutoff date is July 2025.

Core Features

Superior Programming Capabilities: Positioned by the company as “the smartest model for complex agents and coding,” it excels in the SWE-bench Verified test—a benchmark for real-world software engineering capabilities—achieving an accuracy rate of 77.2% in single-model configurations, which rises to 82.0% when parallel inference optimization is enabled, significantly outperforming similar competitors. It supports continuous autonomous work for over 30 hours and can generate approximately 11,000 lines of code at once, covering the entire workflow from project development and bug debugging to code refactoring, making it well-suited for enterprise-level software development needs.

Advantages in Complex Agent Development: It is “the most powerful tool for building complex agents,” equipped with robust planning and coordination, memory management, and sub-agent scheduling capabilities. Paired with Anthropic’s open Claude Agent SDK, developers can easily build sophisticated AI agents featuring permission systems and task decomposition functions, dramatically lowering the barrier to entry for developing high-end agent applications and making it ideal for enterprise-level agent-building scenarios.

High-Efficiency Computer Operations: In the OSWorld real-computer-task benchmark, it leads comparable models with a score of 61.4%, a substantial improvement over the previous version’s 42.2%. It can autonomously perform operating-system-level tasks such as browser navigation, spreadsheet processing, file management, and data entry, seamlessly interacting with various office and professional software tools, making it suitable for automation of office workflows and operations & maintenance scenarios.

Outstanding Reasoning and Knowledge Processing: It performs exceptionally well in specialized domain tests, scoring 83.4% on the GPQA Diamond graduate-level reasoning test, achieving a perfect score on the AIME 2025 math competition test, and attaining an accuracy rate of 89.1% on multilingual question answering (MMMLU). In professional scenarios such as financial analysis, legal document processing, medical research, and STEM fields, it demonstrates exceptional logical reasoning and knowledge application abilities, serving as an efficient assistive tool in specialized domains.

Multi-modal and Parameter Advantages: It supports both text and image inputs and features comprehensive multilingual processing capabilities. Its maximum single-output size is 64K tokens, with a standard context window of 200K tokens; the beta version can support up to 1M tokens via specific headers, meeting the demands of large-scale tasks such as long-form content generation, multi-document integration and summarization, and large-scale code development.

Technical Highlights

Developer Tool Upgrade: Equipped with Claude Code v2, it introduces a new “checkpoint” feature that enables progress saving and rollback, preventing loss of progress due to operational errors. It provides a native VS Code extension and a brand-new terminal interface, allowing direct code execution and file creation within conversational contexts, greatly simplifying the development process and boosting efficiency.

Advanced Security Framework: Adopting the AI Safety Level 3 (ASL-3) release framework, it incorporates high-precision classifier filters that proactively block high-risk content such as chemical, biological, and radioactive materials. Its ability to defend against prompt injection attacks has significantly improved, with a false-positive rate reduced by more than tenfold compared to the previous version. Additionally, it introduces for the first time mechanism interpretability technology, enhancing security assessment and control.

Cost and Deployment Optimization: It supports flexible deployment across multiple platforms, with an input pricing of $3 per million tokens and an output pricing of $15 per million tokens. An innovative prompt caching feature reduces costs by up to 90%, while batch-processing scenarios save up to 50% of costs. This balance between cost and efficiency, achieved on a high-performance basis, makes it suitable for enterprises of all sizes.

Market Impact

The release of Claude Sonnet 4.5 is seen by the industry as a key milestone in the fields of AI programming and complex agent development. Its three-dimensional core capabilities—“superior programming + complex agent development + high-efficiency computer operations”—mark the entry of AI into a deeper stage of application in production-level scenarios. With its leading performance in specialized tests, it has become the preferred model for enterprise-level software development, complex agent construction, high-end automated office work, and specialized research, poised to drive efficiency innovations across related industries. It is particularly favored by sectors such as finance, high-end software R&D, and biopharmaceuticals, where performance and security are of utmost importance.

Related Review: “Claude Sonnet 4.5 vs. GLM-4.6: A Battle of Chinese and Foreign Large Models in Programming—Has the Winner Been Decided Yet?”