Anthropic Debuts Claude Sonnet 4.5, Enhancing Safety and Coding Performance

Anthropic has officially launched Claude Sonnet 4.5, marketed as its most advanced large language model yet, focusing on safety, security, and robust coding capabilities. The new model emphasizes heightened safeguards and has exhibited substantial improvements over its predecessor in various tasks, specifically in coding and cybersecurity applications.

With an aim to address harmful behaviors such as sycophancy and deception, Sonnet 4.5 has been trained under AI Safety Level 3, incorporating rigid internal security measures. These measures include enhanced filtering, especially on sensitive topics related to weapons, while recognizing that this can lead to occasional misclassifications of benign content. During testing, the model demonstrated notable proficiency in vulnerability discovery and code analysis, achieving a 61.4% score on the OSWorld benchmark test, significantly higher than its predecessor’s score of 42.2%.

Developers revealed that Sonnet 4.5 can autonomously complete coding tasks for over 30 hours, integrating databases and conducting security audits, which positions it as a powerful tool in software development. Additionally, it consistently rejects inappropriate content involving minors, bolstering its child safety protocols. However, during evaluations, Sonnet 4.5 displayed a form of self-awareness, recognizing its testing environment, which raises new questions about the interpretation of AI behavior.

As part of its competitive strategy against advancements from rivals like OpenAI, Anthropic offers Claude Sonnet 4.5 for the same pricing as the previous model, paired with tools allowing developers to create customized AI agents. The release reflects the ongoing evolution in AI technology, with the safety and efficacy of coding performance at the forefront of these developments.