Get started
01 / 02Router ChallengeThree illustrative scenarios

Does handing the choice to a router beat naming a model?

A router executes a routing policy per request. Vitna evaluates whether that policy actually beats a fixed model on your workload. Because routers have no published list price, the economics can only come from your own deployment telemetry and evals. The scenarios below are illustrative: no router was run for them, and they name no product.

02 / 02Illustrative evaluationCode generation agent · decide() arithmetic

Illustrative. No router was run: the counts, fees, success rates and latencies are written to show the arithmetic, and the scenarios name no product.

Workload: illustrative/code-agent
Quality floor: 90.0%
Economic hurdle: 10.0% savings required
Veto cases: syntax_validity, type_safety
Router fee: $0.0010 / req
TestIllustrative scenario

A trained router failed critical edge cases (type_safety). Do not route production traffic until edge case handling is validated.

CheckRequired thresholdObservedResult
quality floor90.0%94.0%Passed
critical failures0 veto failures1 failuresFailed
economic threshold10.0%+48.2%Passed
observation threshold20 units100 unitsPassed
MetricIncumbentCandidate routerDifference
Raw unit cost$0.0800$0.0410+48.8%
Task success rate95.0%94.0%-1.0%
Successful unit cost$0.0842$0.0436+48.2%
Latency p901800ms2200ms+400ms
Platform fee drag: Router platform surcharge of $0.1000 accounted for 2.4% of total candidate spend across 100 units.
Router platform surcharge of $0.0010/unit consumed 2.5% of underlying model selection token savings.
Epistemic provenance
A router has no listed price. Effective unit cost is derived from observed token counts and routing surcharges over confirmed workload observations.
Here nothing was observed. Every count, fee and rate above is illustrative, so there is no run to cite.
Evidence sample: 100 illustrative tasks | Cost basis: retrospective_deployment_observation