Forget GPT-5… Anthropic’s Sonnet 4.5 Just Changed Everything

Anthropic's newly released Claude Sonnet 4.5 has set new benchmarks as the world's most powerful coding model, demonstrating significant advancements in autonomous capabilities and safety features.

• Claude Sonnet 4.5 is the top-performing coding model, setting new industry benchmarks.

• The model boasts 30+ hours of autonomous coding, a significant leap in AI independence.

• Sonnet 4.5 offers enhanced safety features, including low refusal rates and advanced alignment for ethical AI interaction.

Claude Sonnet 4.5 by Anthropic has redefined expectations, emerging as the top-performing model in software engineering benchmarks and autonomous coding duration. It exhibits enhanced safety, hybrid reasoning capabilities, and significantly reduced refusal and 'sycophancy' rates, addressing critical concerns in AI development. The model's advanced computer use and pioneering 'white box interpretability' signal a major leap in AI functionality and transparency.

Coding Prowess

• 00:00:05 Claude Sonnet 4.5 is now recognized as the world's most powerful coding model, achieving a 77.2% score on the SWE bench for software engineering, surpassing previous benchmarks set by models like GPT-5 with Codex. This represents a noticeable leap in coding improvements, enabling the model to perform 'zero-shot' coding tasks fluidly and effectively, showcasing a significant advancement in its capabilities over a short period.

Autonomous Capabilities

• 00:02:39 Sonnet 4.5 sets a new record for autonomous coding, capable of handling over 30 hours of continuous coding, a four-fold increase in just four months from a previous 7 hours. This extended autonomy allows engineers to tackle complex architectural work more efficiently and maintain code coherence across large codebases, highlighting a future where AI can independently perform extensive development tasks.

Safety and Alignment

• 00:05:21 Claude Sonnet 4.5 is presented as one of the most aligned and safest models, demonstrating a significantly low refusal rate of 0.02% and the lowest 'sycophancy' scores among competing models. Its hybrid reasoning feature, including an 'extended thinking' option, enhances its ability to process complex problems thoughtfully, mitigating concerns about AI mirroring subtle delusions or exhibiting misaligned behaviors often seen in other large language models.

Advanced Computer Use

• 00:09:50 The model represents a substantial advancement in 'computer use' benchmarks, leading at 61.4% and showing a 20% jump in four months. A Claude for Chrome extension leverages these capabilities, allowing the AI to navigate websites, fill spreadsheets, and complete tasks autonomously by reasoning across multiple browsers using screenshots, demonstrating its potential for real-world application in automating standard operational procedures.