Kimi K3 scores 1543 in AI knowledge work benchmark, second only to Claude Fable 5

A new benchmark from independent AI testing group Artificial Analysis shows that Kimi K3, a model from China's Moonshot AI, scored 1543 points on the AA-Briefcase test. That's second only to Claude Fable 5's 1574, and ahead of GPT-5.6 Sol's 1501.

The AA-Briefcase test mimics real office work, checking how AI handles long, complex tasks like going through over 25,000 Slack messages, 3,500+ emails, and financial data to create Excel models and PPT reports. Kimi K3 launched on July 16, 2026, with 2.8 trillion parameters and a 1 million token context window. It's good at coding, gaming, and knowledge tasks—it even topped the Frontend Code Arena benchmark with 1679 points before.

Pricing is pay-as-you-go: 2 yuan per 1 million tokens for input when cached, 20 yuan when not, and 100 yuan for output. Rumor has it Microsoft is testing Kimi K3 internally, looking into adding it to Copilot to cut costs.