Fireworks says open Kimi K3 rivals Fable 5, with model routing cutting agent task costs by up to 50x

Fireworks benchmarked open model Kimi K3 against closed model Fable 5 across about 1,030 agentic tasks, including software engineering, terminal operations, coding, multilingual implementation, and legal work. The company says routing between the two models reached 93% accuracy while lowering costs versus using Fable alone, with especially large savings on long agentic loops.

The results reinforce a shift from picking one flagship model to using model routing as the default architecture for AI applications. If open models can handle most routine workloads while closed models cover harder edge cases, enterprises may gain more leverage over AI costs and vendor lock-in.