Claude Fable 5 dominates new complex work benchmark, solving only 3% of tasks
a new benchmark exposing ai struggles with realistic multi-week work shows claude fable 5 leading with just a 3% solve rate, signaling agents aren't ready for autonomous knowledge work.
artificial analysis introduced aa-briefcase, a benchmark designed to simulate realistic multi-week knowledge work using thousands of fragmented files. claude fable 5 achieved the highest rubric pass rate among tested models, yet it fully satisfied all criteria for only 3% of tasks, and on 34% of tasks, no model exceeded a 50% score. the results highlight that while top models handle obvious requirements, they fail at synthesizing details across multiple sources, a critical failure mode for autonomous agents. additionally, costs vary drastically, ranging from $0.04 to over $31 per task, complicating the business case for current ai deployments.