Anthropic says its new AI model “maintained focus” for 30 hours on multistep tasks

On Monday, Anthropic introduced Claude Sonnet 4.5, which the company touts as its most advanced AI language model to date, showcasing enhanced capabilities in coding and computer utilization. Alongside this release, Anthropic also launched Claude Code 2.0, an AI command-line agent specifically designed for developers, and the Claude Agent SDK, a toolkit that enables developers to create their own AI coding agents. The company claims that Sonnet 4.5 has successfully maintained focus on a single project for over 30 continuous hours while tackling intricate, multi-step tasks. However, specific details regarding these tasks have not been disclosed. Historically, AI models have struggled with maintaining coherence over extended periods, often losing track as errors accumulate and context windows reach their limits. In previous instances, Anthropic noted that earlier versions of Claude, like the 4.0 models, managed to play Pokémon for more than 24 hours and could refactor code over a span of seven hours. To appreciate the significance of Sonnet 4.5, one must understand the structure of Anthropic's AI language models. The Claude family includes three sizes: Haiku (the smallest), Sonnet (mid-range), and Opus (the largest). The last update to Haiku occurred in November 2024 (version 3.5), while Sonnet was last updated in May (to version 4.0) and Opus received an update in August (to version 4.1). The model’s size, determined by the number of parameters within its neural network, correlates with its contextual depth and problem-solving abilities. However, larger models tend to be slower and more costly to operate, leading AI companies to strive for an optimal balance between performance and cost efficiency. For several years, Claude Sonnet has effectively filled this niche for Anthropic. Developers have shown a fondness for the Claude Code feature, and the company expresses confidence in Sonnet's latest iteration, declaring, "Claude Sonnet 4.5 is the best coding model in the world. It excels in building complex agents, demonstrates superior performance in using computers, and showcases significant improvements in reasoning and mathematics."