Token spend andmodel latency,per call.

GenAI calls become spans and metrics through the OpenTelemetry SDK, using the standard semantic conventions. Model, operation, token counts and finish reason arrive on every call, so a cost jump has a cause you can point at.

insight / llm observabilityLIVE
tokens by modellast 1 hour
gpt-4o1.42M
18.2k calls · p95 1.68 s
claude-sonnet684k
4.9k calls · p95 1.21 s
text-embedding-391k
1.0k calls · p95 74 ms

Token usage and call duration arrive as histograms, split by model and operation.

Models it reads spans fromTrademarks →
What you get

Included on every plan, including the free tier.

01

Token usage and call duration as histograms, split by model, operation and input against output.

02

Finish reasons as first-class data, so truncation and content filtering are countable rather than anecdotal.

03

Model calls as spans inside the request that made them, next to the database and cache work.

04

Message content off by default, so prompts are not stored unless you decide they should be.

How it gets there

Three steps, then it is reporting.

01Install the OpenTelemetry contrib package for your provider.
02Run the app under auto-instrumentation.
03Model calls arrive with tokens and finish reason attached.
What you can ask

What that looks like as a query.

PROMQL

sum by (gen_ai_request_model) (rate(gen_ai_client_token_usage_sum[1h]))

Which model the token spend goes to.

PROMQL

rate(gen_ai_client_token_usage_sum{gen_ai_token_type="output"}[1h]) / rate(gen_ai_client_token_usage_sum{gen_ai_token_type="input"}[1h])

Output against input, where prompt changes show.

PROMQL

histogram_quantile(0.95, sum by (le, gen_ai_request_model) (rate(gen_ai_client_operation_duration_bucket[5m])))

The p95 a user actually waits.

Questions

What AI teams ask before instrumenting.

Something not here? Book thirty minutes with an engineer who works on the platform. Book a call →

Do we need a separate product for this?
No. GenAI spans leave through the OTLP exporter your SDK already has, into the same agent, and read back next to the rest of your telemetry.
Do you store our prompts?
Only if you switch it on. Message content is off by default and needs an explicit GENAI_CAPTURE_MESSAGE_CONTENT to enable. Prompts carry personal data, credentials people paste and whatever retrieval fetched, so turn it on one service at a time or leave it off.
Why is there no dollar figure?
We store token counts and leave the rate to you, in the query or in a recording rule. Prices differ by provider, tier and contract, and a rate baked into the platform would be stale for everyone. Providers quote per million, so $2.50 per million is 0.0000025 a token.
Which providers are covered?
Any that the OpenTelemetry contrib packages instrument, and any you wrap yourself. The GenAI attributes matter more than the mechanism.

Start on the free tier.

20K active series, 50 GB of logs and 50 GB of traces a month, 90-day retention, and the whole of Insights.