All benchmarks

Model routing benchmarks

Compare the performance of model routers, compound models, and other types of multi-model switching logic.

Auto Router is now powered by Jev Router. Learn how it works →
Routers
Favicon for nvidia
Favicon for openrouter
Favicon for sakana
Favicon for unbiased
6 routers
Auto, Fugu Max, Pareto, Auto Code, and Fugu Ultra v2 plus Switchyard model switching
Models
Favicon for anthropic
Favicon for deepseek
Favicon for google
Favicon for moonshotai
20 models
Route and baseline models only; router model choices are unavailable
Results
166
published router and baseline runs

Last benchmark run

Model Routing Frontier

By default each point is a model router’s quality averaged across benchmarks. Cost and speed stay on the horizontal axis instead of being folded into quality, unlike the Model Routing Index in the matrix. Pick a benchmark to compare quality on it, or change what the horizontal axis measures.

Quality frontier

Recorded cost per evaluated task, log scale. Quality score is accuracy relative to the best router on each benchmark (0–10), averaged across benchmarks. A baseline that beats every router is capped at 10. Hollow points are single model baselines, shown for reference. Baselines that score below the axis are left off; they are still in the results table.

Benchmark matrix

The Model Routing Index combines quality, speed, and cost into a 0–10 score relative to the best results on each benchmark. Explore the Model Routing Frontier.

BetterWorseSingle model baselines are shown for reference and are not ranked.

Model routing results