Kimi K3 Scores 1543 Elo in White-Collar Benchmark, Costs 10x More Than K2.6

According to Artificial Analysis' updated AA-Briefcase benchmark, Kimi K3 achieved 1543 Elo points, trailing only Claude Fable 5's 1574 and surpassing GPT-5.6 Sol's 1501. The benchmark evaluates models on information extraction, document analysis, and deliverable generation across 91 confidential tasks.

However, K3's performance gains came with significantly higher costs: $10.57 per task, approximately 10 times that of Kimi K2.6. The model averaged 83 iterations and generated 120,000 tokens per task, requiring 56.4 minutes to completion—roughly 2.5 times longer than Fable 5.