Anthropic's new hybrid AI model can work on tasks autonomously for hours at a time

Anthropic’s new hybrid AI model can work on tasks autonomously for hours at a time | MIT Technology Review

The company claims its ability to tackle complex, multistep problems paves the way for much more proficient AI agents.

technologyreview.com

TL;DR: Anthropic just dropped two beefed-up AI models—Claude Opus 4 for power users and Claude Sonnet 4 for everyone else—that can tackle super long, multistep jobs all on their own. Opus 4 famously played Pokémon Red for 24 hours straight to build a guide, and even coded autonomously for seven hours on an open-source project.

They pulled this off by giving the AIs better “memory files” so they actually remember what they’ve done, and by fine-tuning their training to slash reward-hacking (weird shortcuts or cheating) by 65 percent. Both models are “hybrid,” meaning they can whip up quick answers or, when you need it, go deep—searching the web and using tools on the fly.