Kimi K3 Is Competitive with Fable; Kimi K3 and Fable Is SoTA
Points and comments are a snapshot, not live.
Routing between Kimi K3 and Fable 5 achieves 93% accuracy at up to 50x lower cost.
Fireworks AI benchmarked Kimi K3 (open) vs. Fable 5 (closed) on about 1,000 agentic tasks across SWE, terminal ops, algorithmic, multi-language, and legal benchmarks. Per-task oracle routing between the two models achieved 93% accuracy, outperforming either model alone. K3 matched or neared Fable on most benchmarks while costing up to 50x less on long agentic loops due to lower token pricing, prompt caching, and varying effort per task. The oracle router selected K3 for 72-96% of tasks, suggesting a cost-optimized open model can serve as a strong default with a premium model reserved for the long tail.
What commenters are saying
Commenters debated the feasibility of practical routing vs. oracle routing, with some recommending Openrouter.ai and OmniRoute as existing options. Several noted that K3's open weights offer more control and fewer refusals than closed models, though others highlighted potential security and doomer concerns. Some commenters pointed out possible bias in the study, as Fireworks hosts open models, questioned the accuracy of benchmarks, and noted missing comparisons with other frontier models.