Wayfinder Router: Local Microsecond Model Routing vs OpenRouter — What It Costs to Route

The Model Routing Problem

Model routing is the practice of sending different requests to different models based on complexity — routing a simple code comment to a cheap model and a complex refactor to a frontier model. Done well, it can cut AI coding costs by 40–80% without meaningful quality loss.

The problem with existing routing solutions is that they add cost and latency. RouteLLM and NotDiamond work by calling a separate classifier model on every request. That classifier call costs tokens — typically $0.10–0.30 per 1,000 routed requests — and adds 50–200ms of latency per hop. For high-volume coding agent workflows, the routing overhead can consume 10–15% of your total token budget.

What Wayfinder Does Differently

Wayfinder Router (GitHub: @workweave/router) runs entirely offline. It analyzes prompt structure — length, presence of headers or lists, code blocks, proof markers — and makes routing decisions in microseconds without calling any model. There is no per-request API cost for the routing decision itself.

Installation: npx @workweave/router. It runs as a local proxy on localhost:8080, accepts requests in OpenAI API format, routes them across Anthropic, OpenAI, Gemini, and OpenRouter-connected models, and passes API keys through without storing them.

By default, Wayfinder uses only structural features (prompt length, format markers). Vocabulary-based routing — which looks at the specific words in a prompt — is disabled by default because it didn't generalize well in their blind tests. You can enable it and tune thresholds on your own data.

Cost Comparison: Routing Overhead

For a high-volume coding agent generating 50,000 requests per month:

OpenRouter itself is not a router — it's a unified API gateway. It forwards every request to whatever model you specify. The routing decision is still yours to make. Wayfinder handles that decision step locally.

The Real Saving: Downstream Token Cost

The routing cost is negligible in any scenario. The meaningful saving is in downstream tokens — how many requests get routed to cheap models vs frontier models.

In a typical coding agent workflow analyzed by Wayfinder's benchmark data, roughly 60–70% of requests are structurally simple: short prompts, no code blocks, clear task completion criteria. These are good candidates for DeepSeek V4 Flash ($0.10/$0.20) or Qwen3 Coder Flash ($0.195/$0.975) instead of Claude Opus 4.8 ($5/$25).

On 50,000 monthly requests with average 2K input / 300 output tokens each, routing 65% to DeepSeek V4 Flash and 35% to Claude Opus 4.8:

Limitations to Know

Wayfinder's structural-only routing has real limits. It cannot read intent — a short prompt might still require frontier reasoning if it's asking for a non-obvious architectural decision. The system can only see prompt format, not prompt meaning. Teams who need semantically-aware routing (e.g. "route all security-related prompts to Claude") need a vocabulary-enabled tier or a hybrid approach.

Wayfinder is also self-hosted only. There is no managed cloud version, no SLA, and no fallback if the local proxy goes down. For production coding agent infrastructure, you need to plan process supervision and health checks.

Verdict

Wayfinder is a genuinely useful tool for developers already running multi-model coding agent setups who want zero-cost routing overhead. It integrates with Claude Code, Codex, and Cursor via the localhost:8080 proxy. The 64% cost saving in the estimate above is directionally realistic, though your actual split between simple and complex requests will vary by workflow.

If you are on a single-model workflow and not yet doing routing at all, Wayfinder is a low-friction starting point. If you need semantically-aware routing or managed infrastructure, look at LiteLLM or OpenRouter with custom routing rules instead.