Kimi K3 Agentic Knowledge Benchmark

AA-Briefcase is Artificial Analysis's private benchmark for agentic knowledge work: a model works through realistic tasks spread across thousands of complex input files and must produce deliverables such as spreadsheets, presentations and UI mock-ups. Each result is combined into one Elo rating from correctness, analytical quality and presentation quality. On it, Moonshot AI's Kimi K3 scores 1543, second only to Claude Fable 5 (1574), ahead of GPT-5.6 Sol (max, 1501) and a +727 jump over Kimi K2.6 (816) (artificialanalysis.ai). The launch-day post had given K3 1547 on the same benchmark.

The strength is in substance rather than polish. K3 passes 51% of the grading rubric, behind only Fable 5 (56%), and its analytical quality Elo of 1754 is comparable to Fable 5's 1744. Its presentation Elo of 1471 trails GPT-5.6 Sol (max, 1660), the model Artificial Analysis singles out for presentation.

The cost is the striking part. K3 averages $10.57 per AA-Briefcase task, about 10 times its predecessor's and among the most expensive on the benchmark, and takes 56.4 minutes per task, roughly 2.5 times Fable 5. Artificial Analysis traces this to how K3 works: about 83 turns and 120k output tokens per task, up from 54 turns and 42k tokens for K2.6, plus its token pricing and slower speeds on Moonshot's own API. That reverses the picture from the Intelligence Index, where the launch post put K3 at $0.94 per task, about half of Opus 4.8; on long agentic tasks the headline says it costs more than Opus 4.8 to run.