Skip to main content
Analytics reads canonical project traces from W&B rather than maintaining a separate request-event database. It helps you verify rollout and estimate the economic effect of replacing a model.

Available views

  • request volume over time;
  • traffic share by routed model;
  • project-version distribution;
  • input and output token volume;
  • provider errors;
  • estimated input, output, and total cost;
  • comparison models that are not currently receiving traffic.
Choose a range from one hour through three months. Long ranges use wider time buckets.

Pricing coverage

Analytics uses reported costs when available and otherwise estimates costs using these sources, in order:
  1. A manual model price configured on the provider or fine-tune.
  2. Serverless Inference pricing for a hosted model or a fine-tune’s base model.
  3. The bundled LiteLLM catalog when a matching model is known.
The page reports what percentage of calls had enough pricing information. Compare models only when their coverage is similar. Cost comparisons keep reported spend unchanged and apply the comparison model’s rates to recorded token counts. List-price estimates preserve recorded prompt-cache usage, using ordinary input pricing when a cache rate is unavailable. Observed rates are the comparison model’s reported costs per token in the selected range. They already include its cache usage, so no additional cache discount applies. If you switch models, token counts and cache hits might change. Responses replayed from the exact-response cache keep their original token counts but are priced at zero, because no provider inference ran.

Use analytics during rollout

After changing routing, verify:
  • the new model receives the intended share;
  • the old model remains at the expected residual or zero weight;
  • error rate does not increase;
  • token volume and finish behavior remain plausible;
  • estimated cost moves in the expected direction.
Quality decisions should still come from evaluations and product metrics, not cost charts alone.
The time_range query accepts relative windows such as 1h, 24h, 5d, 30d, or 12w:
See Get project analytics for the response schema.
Last modified on November 3, 2026