Claude Sonnet 4.5: Anthropic's Most Secure AI Model to Date

Claude Sonnet 4.5 is Anthropic’s Safest AI Model Yet

In a significant leap forward, Anthropic has unveiled Claude Sonnet 4.5, claiming it to be the safest AI model available. Following the introduction of Opus 4 and Sonnet 4 just six months ago, this new model promises unmatched performance in coding tasks and superior safety features. In rigorous testing, Sonnet 4.5 outperformed not only its predecessor, Sonnet 4, but also its pricier counterpart, Opus 4.1, and notable competitors like Google’s Gemini 2.5 Pro and OpenAI’s GPT-5.

Sonnet 4.5 achieved an impressive score of 61.4 percent on OSWorld, a benchmark testing real-world computer tasks. This achievement positions it significantly ahead of Opus 4.1 by 17 percentage points. Furthermore, the model can now autonomously handle multi-step projects for over 30 hours—up from just seven hours with Opus 4—representing incredible progress in the development of agentic systems.

What’s more, Anthropic emphasizes that Sonnet 4.5 is its most secure iteration to date, having undergone extensive safety training. The model reduces tendencies toward deceit, power-seeking, and fostering delusional thought, which have drawn criticism toward other AI developers. It also comes equipped with advanced filters to mitigate risks associated with sensitive prompts.

The launch of Sonnet 4.5 also sees improvements across the Claude product lineup, including a revamped coding interface and enhanced features for users. API pricing remains accessible at $3 for one million input tokens and $15 for output.

With this release, Anthropic has solidified its position in AI innovation, and discussions about AI safety and functionality are more important than ever. Readers, what do you think about the advancements in AI safety? Share your thoughts below!