Calculate LLM Cost by Feature and Model

Event contract

Fields the query expects

Field	Type	Why it exists
timestamp_utc	Timestamp	When the model request completed.
feature	Utf8	Stable product feature name.
model	Utf8	Provider model identifier.
total_tokens	Int64	Input plus output tokens.
estimated_cost_usd	Float64	Request cost calculated from versioned model pricing.
status	Utf8	success or failed.

DataFusion SQL

Copy the query

sql

SELECT
  feature,
  model,
  COUNT(*) AS requests,
  SUM(total_tokens) AS total_tokens,
  SUM(estimated_cost_usd) AS cost_usd,
  SUM(estimated_cost_usd) / NULLIF(COUNT(*), 0) AS cost_per_request_usd
FROM llm_requests
WHERE timestamp_utc >= now() - INTERVAL '30 days'
GROUP BY feature, model
ORDER BY cost_usd DESC;

This recipe is read-only and reviewed against Telemetry's DataFusion query patterns. Sample output is deterministic and synthetic; validate field types, thresholds, and business definitions against your own data.

Query result

Thirty-day model cost by feature

Research reports cost more overall despite far fewer requests, pointing to unit-cost optimization rather than an adoption problem.

research_report

921.4 USD

support_draft

496 USD

document_summary

388.5 USD

feature	model	requests	total_tokens	cost_usd	cost_per_request_usd
research_report	reasoning-large	2,180	18,420,000	921.4	0.42
support_draft	fast-small	32,100	24,800,000	496	0.02
document_summary	balanced-medium	8,450	12,950,000	388.5	0.05

Synthetic example output. Run the query against your own event schema and thresholds before using it for operational decisions.

How the SQL works

1Grouping by both feature and model preserves the context needed to explain model migrations.
2Cost per request distinguishes a high-volume inexpensive workflow from a low-volume expensive one.
3Store the calculated request cost at event time using a versioned price table so historical reports do not change when provider pricing changes.

Edge cases to decide

Do not multiply total tokens by one rate when input, cached input, and output tokens have different prices.
Include failed requests if the provider bills them; otherwise add a billing_status field.
Use account or team identifiers when margins vary materially by customer.