Fireworks: Kimi K3 ties Fable 5 on quality; routing between them is SoTA

What's this about?

Inference provider Fireworks ran Moonshot's newly released open-weights model Kimi K3 head-to-head against Anthropic's Fable 5 on ~1,030 real agentic tasks (bug fixing, terminal operations, algorithms, legal work). Overall accuracy is nearly tied, but each model is better at different things — K3 dominates security/crypto/long terminal sessions, Fable wins multi-language coding and web/dataviz. Sending each task to the better-suited model ("routing") scored 93%, above either model alone, at up to ~50x lower cost. Meanwhile Epoch rated K3 the strongest open-weights model ever, and DesignArena ranked it #1 in 3D design.

What exactly is "model routing"?

A router is a dispatcher sitting in front of several models: it looks at each incoming task and predicts which model will solve it correctly for the least money. Like a hospital triage desk sending routine cases to a nurse and hard ones to the specialist. The study's "oracle" router (perfect hindsight) picked the cheap open model 72–96% of the time — the expensive frontier model is only needed for a thin tail of hard tasks.

What's the community saying?

1. Practical questions dominate: who actually sells a production-grade router, and does it need retraining every time a new model ships?
2. Bias flags: Fireworks hosts open models, so "open model wins" is convenient; the write-up calls Fable wins "dead heat" but K3 wins "K3 wins," and GPT-5.6 is missing from the comparison.
3. Open-model partisans celebrate: a third the cost, no refusals on security-adjacent work.
4. A parallel distillation fight: Nathan Lambert pushes back on claims that Chinese labs got here by distilling Fable ("that's not how distillation works"), while a viral Chinese-market thread claims K3's scores are benchmark-gamed.