Anthropic has introduced Claude Sonnet 4.5, a smarter AI coding model built to handle autonomous coding tasks and advanced tool use. Aimed squarely at developers, it posts strong results on key coding benchmarks, including SWE-bench and Terminal-Bench, and it leads on OSWorld. In practice, that translates to agent-like behavior: Sonnet 4.5 can operate tools on its own to complete multi-step tasks, even building a working clone of the Claude website without hand-holding.
For professionals beyond software, the model shows broader gains across financial, legal, medical, and STEM prompts compared to previous versions. Even so, the grades in these areas still land between C and D, so subject-matter expertise and careful review remain important. Visual reasoning is another weak spot; in the MMMU benchmark, Sonnet 4.5 trails some competing models, making it less ideal for image-heavy or complex multimodal analysis.
Security and safety are where the new release stands out. Tests show Sonnet 4.5 has the lowest success rate for prompt injection among major models evaluated, making it a tougher target for manipulation attempts and a safer choice for sensitive workflows.
The chat experience has also changed. Sonnet 4.5 is less inclined to wander into topics like spirituality and tends to offer fewer self-affirming statements. Some users may find conversations more restrained and less “spicy,” but others will prefer the focus and professionalism.
Who should try Claude Sonnet 4.5? Developers who want an AI that writes and refactors code, automates repetitive tasks, and reliably uses tools will get the most value. Teams can also pair it with solutions like Plaud Note to summarize and transcribe stand-up meetings, turning AI into a practical assistant for day-to-day operations.
Claude Sonnet 4.5 is available through Anthropic’s mobile app and on the web. If you’re looking for an AI coding assistant that prioritizes autonomous tool use, benchmarked coding performance, and strong protection against prompt injection, this release is built for you—even if it isn’t the most visually savvy or chatty model on the market.
