Do not trust the router. Verify it.
Sophisticated AI teams test routing systems with their own benchmarks, token counters, latency probes, failure injection, and security review. Neural Router is built for that scrutiny.
Before routing gets clever, the basics must be exact.
Our trust model starts with a simple rule: the router should never silently make a request worse. If a route costs more, takes longer, changes behavior, or trades quality for another objective, that tradeoff has to be explicit, opted into, and visible on the receipt.
- OpenAI-compatible request and response shapes
- True streaming behavior and abort propagation
- Exact token and cost accounting
- Absolute model pins and allowlists
- No silent prompt mutation
- Tenant-isolated cache and policy state
- No silent fallback below your capability or quality constraints
A receipt for every decision.
A routing receipt turns auto into an inspectable system. Your team can see why a provider won, why another lost, and whether the router respected your constraints. Every receipt carries:
- Request profile
- Active service type or policy
- Required capabilities
- Candidate models and providers
- Exclusions and the reason for each
- Scores for cost, latency, quality, risk, and cache affinity
- Winner and a human-readable route reason
- Fallback or escalation events
- Token usage, cost, latency, and overhead
intent · contract clause extraction · json_schema
meets quality floor · lowest eligible cost
below quality floor
no json_schema support
route reason
precision floor · quality ≥ 0.9
cost
$0.0142
ttft
410 ms
Replay and diff
For evaluated traffic, Neural Router compares routed behavior against a direct-to-model baseline: output, cost, latency, tokens, escalation rate, and quality pass rate. The goal is not to claim savings in the abstract. It is to prove savings on your traffic at your quality floor.
Provider scorecards
We measure endpoints across reachability, authentication, API conformance, performance, correctness, reliability, rate-limit behavior, and stability. Public scorecards publish only with preview windows, appeal flows, and clear methodology.
No silent degradation
Cost optimization is only useful when quality is protected. Policies define when to retry, when to escalate, when to degrade visibly, when to waive routing fees, and when to fail cleanly rather than fake a success.
Put the router under scrutiny.
Send one workload, inspect the receipt, and check the numbers against your own counters. That is the fastest way to trust a routing decision.