Calibrate your savings
Tell us roughly what you're using today and we'll estimate how much Aridel could trim from the bill and how much it could lift the average quality of your answers. The numbers come from per-task benchmark scores — your real mileage depends on your real prompts.
Your estimate
With Aridel, you could save — per month and lift your answer quality by —.
Without routing, every easy prompt goes to a random one of your budget-tier models and every hard prompt to a random one of your premium-tier models — on average that lands at the tier mean. The router picks the per-category winner within each tier instead. That gap is the lift you see above.
Note: your provider mix has no premium-tier models in our catalog, so we used the full catalog's premium pool as the high-tier comparison.
How is this calculated?
Cost savings. For each selected task bucket we expand
into the routing categories the worker actually uses. For each category
we mix 60% of best_model_low (cheap-but-good) and 40% of
best_model (top-tier) — a typical easy/complex distribution
from public benchmark data. Your baseline is the average $/M-tokens
of your selected providers' models in the same easy/complex mix.
The percentage gap is applied to the spend range you picked.
Quality lift — tier-fair, mirrors /v1/insights.
Easy prompts compare your budget-tier average quality against the
router's per-category budget pick. Hard prompts compare your
premium-tier average against the per-category premium pick. The
headline number is weighted 60/40 by the easy/complex split. The
implicit assumption is that without a router you'd pick a
model yourself per request — on average you land at the tier mean,
not the per-category best.
Outside-providers signal. For each task bucket, we check whether any underlying category's best_model lives in a provider you didn't select. The percentage is the share of your selected buckets where the answer is yes — i.e. how often the router would reach for a model you don't currently keep in your toolbox.
This is an estimate, not a quote. Real savings depend on your real token mix, your actual prompt distribution, and any volume discounts you already have.