AI / ARTICLE

TaskRouter Shows 100% Accuracy on Classification at $1.04 per 100K Requests

TaskRouter's classification task page shows a router with 100% accuracy at $1.04 per 100K requests, outperforming three off-the-shelf models on cost and accuracy.

Metrics for the Router

The classification task page on TaskRouter shows a router configuration, with the page title reading “Type in Any Task. Get the Best AI Model for It.” The router classifies each request into a configured taxonomy and requires strict structured output. It is listed as Neurometric task-optimized, hosted, and task-prompted, with a Featured marker.

Average input tokens are 483.7, output tokens 55.4, and p50 latency is 2.57 seconds. Cost per 100K requests is $1.04. Token rates are $0.0100 per million for input and $0.1000 per million for output. Accuracy is marked as 100% Best accuracy. The snapshot is from taskrouter-admin, observed on 9/25/2026.

Off-the-Shelf Model Data

Three off-the-shelf models appear on the page for comparison. GPT-6 Luna has 569 average input tokens, 131 output tokens, p50 latency of 2.22 seconds, cost of $12.24 per 100K requests, and accuracy of 76%. Its token rates are $0.1000 per million for input and $0.5000 per million for output. The LiteLLM snapshot was observed on 11/9/2026.

DeepSeek V4.1 Flash 1017 has 617 average input tokens, 495 output tokens, p50 latency of 2.11 seconds, cost of $97.39 per 100K requests, and accuracy of 84%. Token rates are $0.3750 per million for input and $1.5000 per million for output. The taskrouter-admin snapshot was observed on 11/4/2026.

Claude Opus 5.5 has 938 average input tokens, 169 output tokens, p50 latency of 1.72 seconds, cost of $713 per 100K requests, and accuracy of 85%. Token rates are $4.0000 per million for input and $20.0000 per million for output. The LiteLLM snapshot was observed on 11/9/2026.

Page Actions and Observations

The page includes actions for getting task alerts, testing input, and creating an API route.

The router reports the highest accuracy among the listed options and the lowest cost per 100K requests. Claude Opus 5.5 has the lowest p50 latency, but its cost is more than 685 times the router’s cost. DeepSeek V4.1 Flash 1017 reports 84% accuracy at $97.39 per 100K requests, while GPT-6 Luna reports 76% accuracy at $12.24.

END
Feedback