Kimi K3 vs Fable 5: Open-source model costs up to 50× less at 93% routing accuracy

Summary

Kimi K3, as an open-source model, matches the closed Fable 5 in overall accuracy (~92.5%) but could be up to 50 times cheaper; while routing between them achieves about 93% accuracy, K3 handles most tasks.

Why it matters

The study shows that open models can rival proprietary ones in quality and dramatically reduce costs, making large-scale agency workflows more affordable. Additionally, task routing highlights the strategic advantage of combining multiple models' strengths and shifts AI deployment from single-model dependence to a hybrid approach.

Highlights

• K3 nearly matches Fable's accuracy in the SWE benchmark (92.4% vs 92.6%).
• K3 outperforms Fable on terminal tasks; it single-handedly succeeds on 11 security/cryptography problems where Fable fails.
• K3's inference cost on Fireworks, thanks to token pricing and request caching, can be as low as 1/50 of Fable's.

Critical take

The reported 72%-96% task distribution for K3 relies on oracle routing that assumes perfect model selection; real-world routers may struggle to replicate this efficiency across diverse workloads.