Home dashboards

A recorded view of request outcomes, estimated cost avoided, routing, optimization, policy attribution, and operational health for the traffic you select.

WhereOverviewOverviewOpen in Cloud Score

On this page

What you see

Overview, at the top of the sidebar, opens after sign-in. It summarizes recorded requests in the selected UTC period. The default view covers the last seven days of production traffic; it is a snapshot of that period, not a live animation of individual requests. Default shows five KPI tiles (Total requests, Total spend, Avg provider latency, Estimated savings and Success rate) then Request activity, Model usage and Routing strategy mix, then token, cost, model, user, guardrail, reliability and cache charts. Choose Gateway insights for nine charts on traffic, billable spend, savings, latency, outcomes, retries and tokens, over All traffic (its default), Production or Development. Agent usage shows request, cost and latency figures for production requests, with its own filters for agent, attribution, deployment, model, provider, environment and status. The seven sections below are widgets you can add to a saved dashboard. Saved dashboards appear as pinned tabs; Unpinned opens other saved layouts. System dashboards are read-only. Use Create dashboard to name and arrange up to 24 widgets;Edit changes a saved layout and its tab position. Unsaved edits require a discard choice before closing the builder.

Area	What it shows
`Gateway performance`	Requests, estimated savings, tokens avoided, the optimized share, median latency and cache hit rate against the previous period, with reference versus matched cost over time and estimated savings by mechanism with reference coverage.
`Policy impact`	Recorded policy matches, requests with no match, unavailable historical attribution, and the outcomes observed in the selected cohort.
`Routing activity`	Recorded task-to-model-to-provider routes and request activity in UTC time buckets.
`Model intelligence`	Requests and traffic share, mean cost per request, median latency and completion for each model, ranked by recorded traffic.
`Tokens & cache`	Measured prompt compression where snapshots exist, response-cache hits and misses, and in-flight request reuse.
`Operational health`	Provider request completion and observed response time, plus completed, failed and blocked activity over time.
`Optimization opportunities`	Recorded observations and the share of selected requests they affect. They do not predict savings.

Choose the traffic

Choose 1h, 24h, 7d, 30d, or a custom UTC range. Home shows production traffic, except Gateway insights, which starts on All traffic. The range applies to every widget and the request evidence opened from it. Use the Refresh control to refresh now or choose an automatic interval; automatic updates pause while the browser tab is hidden. The rolling 1h and 24h windows advance with refresh, anchored to the current minute.

The address stores the dashboard, range and refresh interval for bookmarks. A model filter opens Optimizations filtered to that model over its default 30 days; a policy, route or cache filter opens it unfiltered. Additional filters are available there. When no requests match, broaden the range or open Playground to record traffic.

Demo traffic is labeled

Outside Gateway insights, Home excludes synthetic demo requests. The All traffic and Development views on Gateway insights and Gateway health label their source so they cannot be mistaken for production value.

Inspect the evidence

Select Requests, Evidence, Performance or Tokens on a Gateway insights chart, or a number in Gateway performance.
An evidence sheet shows exact values, measurement coverage, calculation notes and individual requests for the captured range. This evidence stays fixed while the sheet is open.
Select a policy, route, model or cache outcome.
Optimizations opens with a model, provider or task type filter applied as a chip. Remove the chip there to return to the broader population. It has no policy, cache or optimization feature filter, so those open it unfiltered.
Inspect an optimization observation.
Open an observation's help for its definition and evidence. Choose Filter traffic to open Optimizations with the filters it can apply, or Inspect to open the corresponding evidence view.
Open Performance for a longer investigation.
The link carries the sheet's range: Performance evidence opens Gateway health, Request evidence opens Observability › Requests, and the other sheets open Optimizations over its default 30 days. For one request's route and receipt, open it from the evidence sheet or Logs.

Read the numbers

Measure	How to read it
`Est. savings`	The priced reference model's estimated cost minus the observed list-price cost of the same covered requests. Routing, compression and reuse share this one ledger; provider cache benefit is separate.
`Reference coverage`	The share of selected requests carrying a recorded reference. A reference alone does not guarantee a priced comparison: pending or missing cost evidence leaves savings unavailable rather than zero.
`Tokens avoided`	Recorded compression and response reuse tokens. If coverage is incomplete, the evidence sheet shows the known subtotal rather than inventing a complete total.
`Applied optimization`	The share of requests with an applied mechanism. A prepared prompt alone is not an optimized prompt sent.
`Median latency`	The middle observed request duration; the 95th percentile and sample count are available in Performance evidence.
`Cache hit rate`	Response-cache hits divided by recorded eligible hits and explicit misses. Requests with no cache outcome are excluded.

Comparisons need evidence

The previous-period comparison uses the same duration and filters. When no comparison can be made, the Gateway performance widget reads No comparison; hover it for the reason, such as an empty or zero previous period. Provider completion describes observed request outcomes; it is not provider uptime. See Costs, savings, and estimates for the accounting terms.

Empty, loading, and error states

The first load shows a structured placeholder. When no requests match, the Dashboard suggests broadening the range or opening Playground. A failed load offers Retry; a failed refresh labels the last successful snapshot as stale. Use Refresh to request a current snapshot. Invalid dates in the custom range form show an inline correction; an invalid custom Home range in the URL falls back to seven days.