K3 is a frontier quality open model at a fraction of the cost. Even bigger is that it complements Fable predictably, which makes it possible to get the highest quality intelligence by routing tasks.

🧭 tl;dr: We ran Kimi K3 (open) against Fable 5 (closed) on ~1,000 agentic tasks finding:

We achieved 93% accuracy with routing between K3 and Fable.

Results were up to ~50X more cost effective than Fable alone on long agentic loops, and consistently lower cost across every use case.

How We Measured

We averaged benchmarks, each aimed at a different kind of work , and ran K3 and Fable 5 through the same harness.…