# See request health, token use, cost, cache state, and latency

> KeepRouter records the operational metadata needed to explain a request—model, route, status, token counts, cost, cache state, and latency—without writing prompt or completion bodies to the request log.

_Last reviewed 2026-08-15 · [Editorial review](https://keeprouter.com/editorial-policy#editorial-team)_

## The request is the unit of evidence

An aggregate chart can show a trend, but a request row explains what happened. KeepRouter's user console exposes request-level metadata and time-bucketed usage so a team can move from “cost increased” to the model, token direction, status, and latency behind the change.

## What is recorded

The request log stores timestamp, requested model, public route, HTTP status, measured input and output tokens, recorded cost, cache state or count when available, and latency. It does not store the prompt or completion body. Account and key identifiers are used for access control and attribution.

## Questions the data can answer

- Which model generated the highest input-token volume this week?
- Did cost grow because of more requests, larger context, or longer output?
- Are failures concentrated on one route or model ID?
- Did a repeated request hit the optional response cache?
- Which requests crossed the latency budget?
- Did a key's observed model use match its allowlist and intended application?

## Questions it cannot answer alone

Gateway metadata does not prove answer quality, upstream root cause, or an SLA. A 200 response can still be semantically poor. A routing health check can be green while one upstream account is rate-limited. Join request metadata with application evaluations, traces, and user outcomes rather than treating the gateway as the whole observability stack.

## Debug without leaking content

Record a correlation ID and safe application metadata on your side. When investigating, use the request time, model, status, token counts, and latency to find the matching event. Reproduce with a synthetic prompt if content is sensitive. Do not copy production prompts into support tickets unless your own policy permits it.

## Turn evidence into controls

Use measured usage to set output caps, key spend limits, application concurrency, and model allowlists. Review the slowest and most expensive representative requests before expanding traffic. Observability creates value when it changes a deployment or budget decision.

Use the [AI gateway latency guide](/blog/ai-gateway-latency-guide) to separate gateway overhead from upstream generation time, then apply the [multi-model cost method](/blog/control-multi-model-api-costs) to reconcile attempts, tokens, cache treatment, and charges.

## Frequently asked questions

### Does KeepRouter log prompts or model answers?

Not in the request log. It stores operational metadata such as model, route, status, tokens, cost, cache state, and latency.

### Can I use the dashboard as an SLA monitor?

No. It is usage and request evidence. The public health endpoint also does not make a real inference call to every upstream model.

### How do I investigate one request?

Correlate the time and request identifier with model, route, status, tokens, cost, and latency, then reproduce safely if needed.

### What should trigger a model review?

Repeated failures, latency-budget breaches, token growth, unexpected model use, or cost distribution changes should all trigger a bounded review.

## Sources reviewed

1. [OpenTelemetry trace concepts](https://opentelemetry.io/docs/concepts/signals/traces/)
2. [KeepRouter security and data handling](https://keeprouter.com/security)

## Related guides

- [security](https://keeprouter.com/security.md)
- [status](https://keeprouter.com/status.md)
- [Model routing](https://keeprouter.com/features/model-routing.md)
- [How to control multi-model API costs with evidence, limits, and ownership](https://keeprouter.com/blog/control-multi-model-api-costs.md)

## Measure before you optimize

Send a bounded request and inspect the exact metadata recorded for it.

[Create a free key](https://keeprouter.com/login?returnTo=%2Fconsole%2Fkeys%3Fmodel%3Dfree) · [Live models and pricing](https://keeprouter.com/models.md)