← /blog
· ACE Engineering#llm-router #scorecard #routing #cost #classifier #latency #eval #transfer #managed-api-stack #market-catalog

LLM Router v2: A Scorecard with Automated Model Market Catalog Integration

An updated 800-prompt out-of-distribution evaluation of ACE's local ONNX LLM router: analyzing feature fusion classification (0.798 macro-F1), 300+ model pricing catalog ingestion, 0.00% under-service risk, and 86.6%-97.9% cost savings.

In our earlier benchmark evaluations, we analyzed how static task classifiers behave under distribution shift, demonstrating that held-out in-distribution F1 scores (~0.975) dropped significantly when tested against an independent, out-of-distribution (OOD) benchmark. Crucially, legacy keyword routing exposed an under-service risk of over 10.1% overall and > 53% on complex math and reasoning queries.

With LLM Router v2, we overhauled the classification engine, feature representation, and market routing policies. By integrating the Automated Model Market Catalog alongside a multi-layer Feature Fusion MLP architecture—combining dense sentence embeddings, TF-IDF n-grams, and 12-dimensional rule vectors—Router v2 achieves 0.7984 macro-F1, 77.0% classification accuracy, and 92.63% route agreement.

When running under Arm A3 (Trained Classifier + Market Matrix), the system achieves $0.00 cost regret, 86.62% dollar savings against a mid-tier gpt-4o baseline, and completely eliminates under-service risk to 0.0000% (0 under-served requests out of 800)—all operating locally on CPU with a median routing latency under 7 ms.


1. 800-Prompt Benchmark Evaluation Summary

We evaluated four router execution arms against an independent, 800-prompt disjoint evaluation corpus (data/router_eval/router_eval.jsonl):

+-------------------------------------------------------------------------------------------------------------------------+
| LLM ROUTER V2 SCORECARD (800-PROMPT OOD BENCHMARK)                                                                      |
+-------------------------------+-----------------------+---------------------+-------------------+-----------------------+
| Router Execution Arm          | Macro-F1 Score        | Classification Acc  | Route Agreement   | Under-Service Rate    |
+-------------------------------+-----------------------+---------------------+-------------------+-----------------------+
| A0: Keyword (Legacy Deployed) | 0.6552                | 60.38%              | 87.25%            | 10.13% (81 / 800)     |
| A1: Centroid (Embedder)       | 0.7124                | 70.37%              | 88.38%            | 8.13%  (65 / 800)     |
| A2: Trained (Hybrid Head)     | 0.7984                | 77.00%              | 89.88%            | 0.25%  (2 / 800)      |
| A3: Trained + Market Catalog  | 0.7984                | 77.00%              | 92.63%            | 0.0000% (0 / 800)     |
+-------------------------------+-----------------------+---------------------+-------------------+-----------------------+

2. Ingestion of the 300+ Market Model Pricing Catalog

The integration of our 300+ model pricing market catalog transforms the router from a static tier assigner into a real-time market cost optimizer.

Incoming Request
  │
  ├── 1. Structural Filter ──► Image / >32k Tokens / JSON ──► Direct Fast-Path Routing
  │
  ├── 2. Decisive Rule Precedence ──► High-Precision Regex (Math, Code, Reasoning)
  │
  ├── 3. Local Feature Fusion MLP  ──► Dense Embedding (bge-small 384-dim)
  │                                    ├── Sparse TF-IDF N-Grams
  │                                    ├── 12-Dim Rule Vector Features
  │                                    └── Multi-Layer Perceptron (GELU/ReLU)
  │
  └── 4. 300+ Market Catalog ──► Evaluates 300+ Market Models (OpenAI, Anthropic, Gemini,
                                    DeepSeek R1/V3, Llama 3.3/3.1, Qwen 2.5, Perplexity)
                                    ──► Serves Optimal Pareto Candidate

Key Performance Gains from 300+ Market Model Ingestion

Performance Dimension Legacy Keyword Router (A0) Router v2 + 300+ Market Catalog (A3) Operational Impact
Under-Service Rate 10.13% (81 failures) 0.0000% (0 failures) 100% Elimination of Task Quality Regressions
Route Agreement w/ Gold Oracle 87.25% 92.63% +5.38% Increase in Intent Alignment
Reasoning Task Under-Service 53.66% 0.00% Zero Failures on Complex Reasoning
Math Task Under-Service 44.58% 0.00% Zero Failures on Mathematical Prompts
Cost Regret ($/1M Requests) -$310.02 USD $0.00 USD Eliminates Fake Savings from Under-Servicing
Cost Savings vs. GPT-4o Baseline 95.71% 86.62% Saves $23,976 USD per 1M Requests

3. Financial & FinOps Impact

Evaluating 1,000,000 requests against a standard mid-tier gpt-4o baseline demonstrates dramatic cost savings without compromising task quality:

  • Flagship Baseline (all-claude-opus-4-1): $180,139.35 USD / 1M Requests
  • Mid-Tier Baseline (all-gpt-4o): $27,681.29 USD / 1M Requests
  • Router v2 + 300+ Market Catalog (Arm A3): $3,704.90 USD / 1M Requests
  • Net FinOps Savings: 86.62% Cost Reduction ($23,976.39 USD saved / 1M requests)

Get Started in Minutes

Ready to deploy local LLM routing with 0.00% under-service risk and 86.6% cost reduction?

Mint your developer API key and start building today:

👉 Get Started & Mint Your Developer Key