TLDR: Anthropic has launched Claude Sonnet 4.5, an advanced AI model it claims is the best globally for coding, agent development, and computer interaction. The new model showcases significant performance improvements in software engineering benchmarks and introduces features like checkpoints, enhanced context editing, and a dedicated SDK for agent creation.
San Francisco, CA – September 29, 2025 – AI research company Anthropic today announced the release of Claude Sonnet 4.5, its latest artificial intelligence model, which the company boldly positions as the ‘best coding model in the world’ and the ‘strongest model for building complex agents.’ This new iteration of Claude Sonnet also claims superiority in general computer usage, marking a significant leap in AI capabilities.
According to Anthropic’s internal evaluations, Claude Sonnet 4.5 achieved an impressive 77.2% accuracy rate on the SWE-bench Verified benchmark, a widely recognized test for an AI model’s ability to solve real-world software engineering tasks. This performance reportedly surpasses competitors, including OpenAI’s GPT-5 Codex, which scored 74.5%, and Google’s Gemini 2.5 Pro, at 67.2%. Furthermore, Sonnet 4.5 demonstrated its prowess in agentic terminal coding and agentic tool use, outperforming its rivals in these critical areas. It also showed strong capabilities in financial analysis, where it again outpaced GPT-5 and Gemini 2.5 Pro. However, the model scored slightly below these competitors in high school math (without tools), graduate-level reasoning, and visual reasoning.
Sonnet 4.5 is designed to excel in powering agents for diverse applications such as financial analysis, cybersecurity, and research, capable of coordinating multiple agents and processing high volumes of data with enhanced reliability. The model also boasts superior instruction following, tool selection, error correction, and advanced reasoning, making it suitable for customer-facing AI assistants and complex AI workflows.
Key new features accompanying the Claude Sonnet 4.5 release include:
Checkpoints in Claude Code: This allows users to save their progress during coding tasks and revert to previous states, offering greater flexibility and control.
Enhanced API Tools: A new context editing feature and a memory tool have been integrated into the Claude API, improving long-running task management.
In-App Code Execution and File Creation: Claude apps now support direct code execution and the creation of various file types, including spreadsheets, slides, and documents, directly within the chat interface.
Claude for Chrome Extension: Max users can now access a dedicated Chrome extension.
Claude Agent SDK: Anthropic has made a Software Development Kit available, empowering developers to build custom AI agents leveraging Sonnet 4.5’s advanced capabilities.
‘Imagine with Claude’: An experimental preview for Max subscribers that enables the creation of ‘software on the fly’ by responding in real-time to user requests to build programs.
Extended Thinking Mode: This feature allows the model to engage in deeper reasoning for complex tasks, multi-step coding projects, and in-depth research, prioritizing accuracy over latency.
Michele Catasta, President of Replit, lauded Sonnet 4.5’s editing capabilities, stating, “Claude Sonnet 4.5’s edit capabilities are exceptional – we went from a 9% error rate on Sonnet 4 to 0% on our internal code editing benchmark. Higher tool success at lower cost is a major leap for agentic coding.”
Anthropic has maintained the pricing for Sonnet 4.5 through the developer API at $3 per million input tokens and $15 per million output tokens, consistent with its predecessor, Sonnet 4. While this makes it significantly more affordable than Claude Opus, it remains more expensive than OpenAI’s GPT-5 and GPT-5 Codex. The company emphasizes Sonnet 4.5’s utility beyond coding, positioning it as a versatile chatbot for various workplace tasks, including financial services, cybersecurity, and legal applications, differentiating it from models often used for non-work-related conversations.
Also Read:
- Anthropic Intensifies Global AI Race with Strategic Entry into India
- Gemini 2.5 Flash-Lite Preview Crowned Fastest Proprietary Model with 50% Output Token Reduction
Furthermore, Sonnet 4.5’s ability to directly execute code in a sandboxed server environment, supporting Python and Node.js, and its capacity to clone GitHub repositories and install software packages from NPM and PyPI, underscore its robust functionality for developers.
