Preference-aware routing for Claude Code 2.0 (www.archgw.com)

Anthropic’s Claude Code 2.0 is introducing preference-aware routing via Arch-Router, a practical system that lets developers encode human-readable routing policies to map coding tasks to the best-fit LLMs. Instead of forcing every workflow through one model or relying solely on benchmark rankings, teams describe task-level preferences (e.g., “code generation” vs “code review”) in a YAML config and assign providers/models accordingly—OpenAI, Anthropic, GPT‑4o, Claude Sonnet, or smaller models for low-latency work. The release includes a Dockerized demo on GitHub that shows how to edit config.yaml, spin up an Arch Gateway, and monitor routing decisions (logs and a pretty_model_resolution.sh utility reveal which model handled each Claude Code 2.0 request). This matters because it decouples “what” the task is from “how” it’s executed, allowing real-world evaluation criteria—domain fit, latency, cost, reliability—to drive routing rather than synthetic benchmarks. Technically, Arch-Router implements policy-driven model selection at request time, supports multi-provider setups, and enables live reconfiguration so developers can tune routes (e.g., route heavy reasoning to a larger model, generation to a smaller one) without brittle orchestration code. For ML teams and tool-builders, this simplifies building more efficient, accurate, and cost-effective agentic pipelines.