Every agent step deserves its own model. Seracade routes to the one that fits, continuously.
Per step classification, empirical scoring on your own traffic, frontier optimal routing up or down. One base URL change. 15% of the price difference per routed call, free until $500 per month.
No. Seracade logs model names, latency, and token counts, not prompt content. During the audit, a structural hash of each prompt is used to classify task types. Raw text is never stored or sent to third parties.
What if I already use multiple models?
Seracade audits the full distribution across your calls and evaluates whether each routing decision is optimal. Where your current model selection is already correct, it confirms that. Where it finds a better option, it surfaces it.
How does Seracade measure output quality?
Seracade replays each task type against candidate models and scores outputs on semantic similarity, completeness, and task-specific criteria. Scores are normalized against your current model as the baseline, so quality comparisons are always relative to what you already ship.
What models are on the frontier today?
The frontier updates continuously as new pricing and benchmarks publish. Today's frontier: Gemini 2.0 Flash, GPT-4.1-mini, DeepSeek-R1, Gemini 2.5 Pro, Claude Sonnet 4.6, Claude Opus 4.6. When a new model enters the market, Seracade benchmarks it within 24 hours. Should you move your default model to one of these? Yes, and we encourage it. But no single default covers every task classification. Each task type has a different optimal model. Seracade routes each call to the right frontier model for that task type automatically.
How is the 15% fee calculated?
For each routed call, Seracade takes 15% of the absolute price difference between your original model and the routed model. Example: 100K calls/month, 65% routed, average price difference of $0.0018 per call. Routing value: $117/mo. Seracade fee: $17.55/mo. Net to you: $99.45/mo. Billing starts only after routing value exceeds $500 in a calendar month.
You set one model for all tasks. Your tasks don't agree.