Cost & Performance Metrics
These two views sit between the prebuilt dashboards and the AI Optimizer. They line up cost, tokens, latency and quality scores so you can see where a model swap or a prompt change is worth testing.
Open them from Gauge AI → Cost Performance Metrics and Gauge AI → Latency & Analytics. Both need Gauge AI read access (or the Cost & Latency dashboards sub-module): Owner, Department Admin, Security & Compliance and Cost Analyst have it.
Both pages have a date range picker (limited by your plan's data access period), an Environment selector and a filter builder.
Cost Performance Metrics
| Section | What it shows |
|---|---|
| Traces | Total traces tracked and the trace names with the most traces. |
| Model costs | Per model: tokens and cost in USD, with the total cost for the range. |
| Scores | Per score name and source: count, average, and how many values were 0 and 1. |
| Traces by time | Trace counts, and observations by level, over time. |
| Model Usage | Tabs for Cost by model, Cost by type (input, output and other usage types), Units by model and Units by type. |
| User consumption | Tabs for Token cost and Count of Traces per user. Users appear when traces carry a userId. |
| Scores chart | Moving average per score over time. |
Latency & Analytics
| Section | What it shows |
|---|---|
| Trace latency percentiles | p50, p90, p95 and p99 per trace name. |
| Generation latency percentiles | p50, p90, p95 and p99 per generation name. |
| Span latency percentiles | p50, p90, p95 and p99 per span name. |
| Model latencies | Generation latency per model over time, with tabs for the 50th, 75th, 90th, 95th and 99th percentile. |
| Scores Analytics | Pick one or more scores to see their distribution and averages over time, based on up to 10,000 scores. |
Reading the trade-off
The point of these views is comparison:
- High cost, similar scores. A model in Model costs that costs much more than another without better scores is a candidate for the AI Optimizer.
- Slow model. A model with a high p95 in Model latencies is a candidate for a latency optimization run (real-time replay only).
- Expensive users or trace names. User consumption and the Cost Dashboard show who and what drives spend.
Scores come from LLM-as-a-Judge evaluators, annotations or scores you send with the SDK or API. Without scores, the quality side of the comparison stays empty.
Next steps
- AI Optimizer: turn a hotspot into a tested model recommendation.
- Scores: how quality numbers are produced.
- Users: per-user cost and usage under Gauge AI → Users.