Requesty Adds Jev-Powered Model Routing

Jev-powered routing is a practical step toward model choice becoming application logic, but its savings depend heavily on task-specific calibration.

• –Jev selects among allowlisted model branches using structured decisions rather than free-form generated text.
• –Requesty’s internal 129-task benchmark matched Opus-level results at roughly 71–72% of the cost.
• –The router runs once per session, reducing decision overhead and avoiding model hopping mid-task.
• –Results did not transfer cleanly to generic SWE-bench or DeepSWE routing, so developers need held-out evaluation on their own workloads.
• –A strong default-model fallback and confidence threshold are essential when the router misclassifies a hard task.