Skip to content
Lookspan

Pricing & cost tracking

Lookspan tracks token usage and computes dollar cost for every span and trace — server-side, so your agent never has to do the math.

How cost is computed

The collector aggregates input/output/cached/reasoning tokens per span and computes cost_usd from a built-in model pricing table, keyed by model name. Costs roll up from spans to the trace and feed the Costs & overview view (total, by model, by provider, by agent, and cost per day).

Because the math is server-side, sending costUsd: 0 from your agent is fine — Lookspan fills it in. The OTel and SDK paths both rely on this: token counts + model name in, dollar cost out.

Inspect cost

Terminal window
# Cost breakdown — total, by model, provider, agent
curl localhost:3100/api/costs/summary


# Stats summary — totals, error rate, latency p50/p95/p99, cost per day
curl localhost:3100/api/stats

In the dashboard, recent traces show per-row cost mini-bars, and the overview page charts cost per day alongside latency and error rate.

Keep prices current

Model prices change. Override the built-in table with your own JSON file:

Terminal window
npx lookspan --pricing ./pricing.json
# ...or LOOKSPAN_PRICING=./pricing.json npx lookspan

The custom table is merged over the built-in one, so you only need to specify the models you want to correct or add.

Each entry needs model, inputPer1M, and outputPer1M; cachedInputPer1M and reasoningPer1M are optional. Cached input and reasoning tokens are treated as subsets of input and output respectively, so they are billed at their own rate without being double-charged. When reasoningPer1M is omitted, reasoning tokens are billed at the output rate.

Reasoning cost, itemized

Reasoning tokens are billed correctly (at their own rate when set), but they are also itemized so you can see what the reasoning premium actually costs — per span and per model, not just blended into output. cost/summary returns:

  • reasoning — total USD attributable to reasoning tokens (a sub-component of total, never an extra charge).
  • reasoningByModel — that cost broken down by model: the routing dimension, so a “use a smarter model on harder requests” strategy is legible per request type instead of only in aggregate.

The figure is computed even when a model has no explicit reasoningPer1M: it is then the share of the output bill those tokens account for, which is exactly the signal that tells you whether setting a dedicated reasoning rate is worth it. In the dashboard, the overview shows reasoning as a share of total cost and charts Reasoning cost by model when any is present.

Cost-based alerts

Pair cost tracking with an alert to get notified when a single trace gets expensive:

Terminal window
npx lookspan --alert-cost 1.00   # toast + notification when a trace tops $1