Kimi K3 matches Fable 5 quality at 50x lower cost via smart routing
Open model K3 rivals closed Fable; routing yields 93% accuracy and huge savings.
Kimi has introduced K3, an open-weight model that matches the frontier quality of closed rival Fable 5 across hundreds of agentic tasks. Benchmarked on ~1,030 real-world tasks—including SWE-bug fixes, terminal ops, algorithms, multi-language coding, and legal work—K3 and Fable scored within a few points of each other overall (SWE: 92.4% vs 92.6%). However, K3 shines on symbolic math, dev tooling, security, and crypto, while Fable excels at web/data viz and broad language support. The real breakthrough is cost: by routing tasks between K3 and Fable, Kimi achieved 93% accuracy with up to 50x cost savings on long agentic loops, as K3's token pricing and caching make it dramatically cheaper for most workloads.
K3's cost advantage is driven by lower token pricing and prompt caching, though it uses more tokens on some tasks (e.g., SWE: 55 turns, 1.3M tokens vs Fable's 21 turns, 130K). On terminal tasks, the opposite occurs: Fable spirals to 64 turns and 1.5M tokens. This complementary behavior means a practical router can predict which model to use for maximum efficiency. Kimi also announced a Series D funding round and $1B ARR, signaling strong commercial momentum. The result is a practical path for enterprises to deploy high-quality AI agents without paying for a single model's inefficiencies.
- K3 (open) and Fable 5 (closed) score nearly identical on SWE-bench (92.4% vs 92.6%), but specialize in different domains.
- Routing tasks achieves 93% accuracy and up to 50x cost savings on long agentic loops compared to using Fable alone.
- K3 is cost-optimized for most work types due to lower token pricing and prompt caching, despite sometimes using more tokens.
Why It Matters
Enables cost-effective AI agents by routing tasks between open and closed models for optimal performance.