Anthropic’s AI Sustains 30-Hour Focus on Complex Tasks
▼ Summary
– Anthropic released Claude Sonnet 4.5, its most capable AI model to date, alongside Claude Code 2.0 and the Claude Agent SDK for developers.
– The new Sonnet 4.5 model can work continuously for over 30 hours on complex tasks, addressing past issues where models lost coherence over time.
– Anthropic’s Claude family includes three model sizes: Haiku (smallest), Sonnet (mid-range), and Opus (largest), with Sonnet balancing performance and cost.
– Larger models generally have greater contextual depth and problem-solving ability but are slower and more expensive, making mid-range models like Sonnet ideal for trade-offs.
– Anthropic claims Claude Sonnet 4.5 is the world’s best coding model, excelling in building complex agents, computer use, reasoning, and math.
Anthropic has launched its newest AI language model, Claude Sonnet 4.5, describing it as the most advanced system the company has ever created. This release comes with significant enhancements in coding proficiency and computer interaction skills. Alongside the core model, Anthropic introduced Claude Code 2.0, a command-line AI agent aimed at developers, and the Claude Agent SDK, a toolkit enabling programmers to construct their own AI coding assistants.
A remarkable feature of Sonnet 4.5 is its ability to maintain focus on a single, intricate project for durations exceeding 30 hours. This represents a substantial leap in persistent task execution for AI agents, which have historically struggled with coherence over extended periods. Earlier models tended to degrade as mistakes piled up and their short-term memory, known as the context window, became saturated. Previously, Anthropic demonstrated that Claude 4.0 could play Pokémon for more than a day or refactor code for seven hours straight, but the new model pushes these boundaries considerably further.
To appreciate the significance of Sonnet, it helps to understand Anthropic’s model lineup. The company typically offers three distinct model sizes: Haiku, the most compact; Sonnet, the mid-tier option; and Opus, the largest and most powerful. These models receive periodic updates, with Haiku last upgraded to version 3.5 in November 2024, Sonnet moving to 4.0 in May, and Opus reaching 4.1 in August. The number of parameters, essentially the neural network’s stored values, generally correlates with a model’s contextual understanding and problem-solving prowess. However, larger models also demand more computational resources, resulting in slower speeds and higher operational costs. For years, Claude Sonnet has successfully occupied that middle ground, delivering an optimal balance of performance and affordability.
Many software engineers have grown fond of Claude, particularly through its Claude Code feature. Anthropic expresses strong confidence in the coding capabilities of this latest iteration. On its official website, the company states, “Claude Sonnet 4.5 is the best coding model in the world.” They further claim it is the superior choice for developing sophisticated agents, excels at computer utilization, and demonstrates major improvements in logical reasoning and mathematical tasks.
(Source: Ars Technica)
