- K3 is an open, low-cost frontier model that complements Fable predictably, enabling high-quality intelligence via task routing.
- Oracle routing shows K3 selected for 72–96% of tasks, suggesting a near-perfect router is achievable with more data.
- Benchmarking over ~1,030 tasks across five families (SWE, Terminal, Algorithmic, Multi-Language, Legal) shows near-tie in overall quality.
- Models have discrete strengths: K3 excels in symbolic math, dev tooling, and terminal tasks; Fable leads in web, data viz, Java/Python/C++.
- K3 is far cheaper due to token pricing, prompt caching, and variable effort per task, though sometimes slower.
- Per-task routing consistently outperforms any single model, achieving higher quality at lower cost.
- The best AI is a mixture of models from different labs, not a single provider.