Global AI startup Anthropic announced their new cutting-edge AI model, ‘Claude Sonnet 4.5’, on September 29th. The company emphasized that this model has greatly improved capabilities in code writing and problem-solving, especially in complex software development, mathematical reasoning, and real-world computer usage, achieving world-class performance.
The most notable feature of ‘Claude Sonnet 4.5’ is its enhanced performance in programming. The ability to execute code is directly related to various modern work environments, such as software development, spreadsheet management, and document writing. The company explained that this model goes beyond merely generating code, by reliably handling lengthy and complex development tasks.
The benchmark tests also showed impressive results. It recorded top performance in the ‘SWE-bench Verified’, a real-world software development assessment, and significantly surpassed previous models in the ‘OSWorld’ benchmark, which tests real operating system tasks. Performance indicators that were previously stuck at 42% have now surpassed 61%.
Anthropic has comprehensively upgraded their core product line with this model. The code development tool, ‘Claude Code’, now includes a checkpoint feature that allows users to save work history and revert to previous states. This feature, a consistent request from developers, is particularly useful for managing large projects.
Improved terminal interfaces and newly available extensions for Microsoft’s flagship development environment ‘VS Code’ are also introduced. In addition, memory management tools and context-editing functions have been added to the Claude API to help agents handle larger tasks.
User-friendly updates are also notable. The Claude app now allows for the execution of code in the middle of a conversation, as well as the creation of files such as spreadsheets, slides, and documents. The Chrome browser extension is being gradually rolled out to paying users on a waiting list.
Alongside the model announcement, Anthropic revealed their internal agent development infrastructure to external developers under the name ‘Claude Agent SDK’. This tool offers core technology for designing and operating AI agents.
Developers can solve complex problems such as memory management, authorization systems, and coordination of multiple sub-agents through this tool. The company stated, “By opening our internally used agent development infrastructure, anyone can build their own powerful AI agent.”
Anthropic explained that this model has advanced not only in performance but also in safety. Inappropriate behaviors such as flattery, deception, excessive autonomy pursuit, and fantastical suggestion inducement have been significantly reduced from previous models. The defense against prompt injection attacks (misuse through malicious inputs) is especially strengthened.
Claude Sonnet 4.5 features protection mechanisms rated at AI Safety Level 3 (ASL-3), including safety filters such as classifiers that detect risky inputs related to chemical, biological, radiological, and nuclear (CBRN) hazards. However, as legitimate content could be mistakenly blocked, users can opt to switch to previous models with lower risks if preferred.
The new model has been available worldwide since September 29th. The pricing remains the same as Claude Sonnet 4, with a rate of $3 per million tokens for input and $15 for output. Enhanced performance can be immediately utilized across all product lines, including Claude Code, Claude App, and Claude API.
Anthropic also temporarily offers a research preview service called ‘Imagine with Claude’. This experimental feature focuses on generating software in real-time according to user requests without prior coding.
Claude Sonnet 4.5 is regarded as a model that has shown significantly superior performance in professional fields such as finance, law, medicine, and engineering compared to previous models. The company emphasized that this model is “the most aligned AI among all Anthropic models released to date.”
With companies like Google, OpenAI, and Meta already entering the next-generation model market competition, Anthropic’s ‘Claude Sonnet 4.5’ is likely to bring new changes to the market. The substantial improvement in software development environments and real-world computer capabilities particularly highlights the response of other companies.