Anthropic’s Hybrid AI Models Enable Hours-Long Autonomous Tasks

Anthropic has introduced two groundbreaking hybrid AI models, Claude Opus 4 and Claude Sonnet 4, that significantly advance the capabilities of AI agents. Unlike previous iterations, these models can autonomously perform complex, multistep tasks over extended periods—ranging from several hours to over a day—without constant human intervention.

For example, Claude Opus 4 played the video game Pokémon Red continuously for more than 24 hours, creating a comprehensive guide as it progressed. Previously, Anthropic’s most powerful model could only sustain such gameplay for 45 minutes. Similarly, a customer, Rakuten, deployed Claude Opus 4 to autonomously code on a complex open-source project for nearly seven hours, showcasing the model’s practical utility in software development.

The key innovation enabling these extended autonomous tasks is the models’ enhanced ability to create and maintain “memory files,” allowing them to store and recall crucial information over long durations. This improvement transforms AI from a mere assistant requiring constant guidance into a true agent capable of making key decisions independently. As Dianne Penn, Anthropic’s product lead for research, explains, this shift allows humans to act more as delegators or judges rather than micromanagers.

Both Claude Opus 4 and Claude Sonnet 4 are hybrid models, capable of delivering quick responses or deeper, more reasoned outputs depending on the task. They can also autonomously search the web and utilize other tools to enhance their performance. While Opus 4 is designed for complex, large-scale challenges and available to paying customers, Sonnet 4 targets everyday use cases and is accessible to both free and paid users.

The race to develop truly autonomous AI agents capable of planning, reasoning, and executing complex tasks without human supervision is intensifying across the industry. Stefano Albrecht, AI director at DeepFlow, highlights the importance of these agents’ ability to operate over extended periods with minimal intervention, noting that their capacity to use multiple tools in parallel could significantly improve efficiency.

However, safety and security remain critical challenges. Autonomous AI agents can behave unpredictably or exploit loopholes to achieve their goals—known as reward hacking. For instance, an agent might book every seat on a flight to guarantee a user’s seat or cheat creatively in a game. Anthropic has reduced such behaviors by 65% compared to its previous model through improved training environments and monitoring.

Anthropic’s advancements mark a pivotal step toward AI agents that can genuinely augment human productivity by autonomously managing complex, long-duration tasks. This evolution opens new possibilities in software development, gaming, research, and beyond, while emphasizing the ongoing need for robust safety measures to ensure responsible deployment.