OpenRouter

OpenRouter's Broadcast sends a trace for every API request you make through OpenRouter. Point it at simpledatalogs and each LLM call becomes one llm_call event, with the model, provider, tokens, cost in USD and latency as properties. Your app's code doesn't change.

Prompts and completions are never stored.

Set it up

1. In simpledatalogs, create a secret key (sdl_sec_…) on the deployment the calls should count under, such as a new deployment named openrouter.
2. In OpenRouter, open Settings → Observability and turn on Enable Broadcast. In an organization, you need the admin role.
3. Click the edit icon next to Webhook and fill in:

Field	Value
URL	https://edge.simpledatalogs.com/openrouter/v1/traces
Method	POST
Headers	{"Authorization": "Bearer sdl_sec_…"}

4. Under Privacy, turn on Privacy Mode. simpledatalogs drops prompts and completions either way, and with Privacy Mode they never leave OpenRouter.

Keep Privacy Mode on

A trace over 512 KB is refused and lost. Long prompts and agent transcripts get there quickly; without them, a trace is a few KB.

5. Optional: under Additional generation metadata, turn on Cost to add the billed tier, BYOK and estimate flags. Costs and tokens come without it.
6. Click Test Connection, then save. OpenRouter saves only when the test passes.
7. Make one request through OpenRouter. An llm_call event shows up in Logs within a minute.

OpenRouter's test trace (a made-up GPT-4 Turbo call costing $0.02) arrives as llm_call_test instead, so it never adds to your llm_call spend.

OpenTelemetry Collector works too

OpenRouter's OpenTelemetry Collector destination sends the same traces. Use the same URL as its Endpoint and the same headers.

To send only some keys' calls, pick them under the destination's API keys, or exclude the ones to leave out. Sampling under 1.0 sends fewer traces, so counts and spend drop with it.

What's stored

One llm_call event per generation. Provider retries and fallbacks inside one call don't count twice.

Property	What
model	The model that answered, such as openai/gpt-4o-2024-08-06
requested_model	The model you asked for, only when routing picked another
provider, provider_slug	Who served it, such as OpenAI and openai
input_tokens, output_tokens, total_tokens	Token counts
cached_tokens, cache_write_tokens, reasoning_tokens	When the provider reports them
cost_usd	What OpenRouter charged for the call, in USD, fees included
input_cost_usd, output_cost_usd	The prompt and completion parts of the cost
input_unit_price_usd, output_unit_price_usd	The price per token OpenRouter charged
duration_ms, ttft_ms, inter_token_latency_ms	Total time, time to first token, and the mean gap between tokens
status, error, finish_reason	ok or error, the error message, and why the answer ended (stop, length …)
generation_id	OpenRouter's generation id, to look the call up in OpenRouter
api_key_name	Which OpenRouter key made the call
service_tier, is_byok, usage_is_estimated, cancelled	With Cost on: the tier it was billed at, whether it used your own provider key, whether OpenRouter had to estimate usage, and whether the call was cancelled
span_name, trace_id, span_id	The trace's names and ids
gateway	Always openrouter

• A failed call is an error-level event, so it shows in error counts and alerts.
• The request's user becomes the event's user, and session_id its session. An email as user is pseudonymized like any email user id.
• Flat values in the request's trace object ("trace": {"feature": "support"}) become properties too, up to 20.
• Prompts, completions, messages, tool definitions and the provider's raw usage are always dropped.
• A resent trace lands on the same event, so nothing counts twice.

Chart your spend

Add a board with these widgets:

• Number, event llm_call, measure sum of cost_usd: what you spent in the range.
• Stacked bars, event llm_call, split by model (or provider): calls per model over time.
• Number, measure avg of duration_ms: how slow calls are.
• Number, measure rate, condition status is error: the share of calls that failed.

The Spend tab prices calls by a fixed unit cost. OpenRouter already sends the real cost, so chart cost_usd on a board instead.

Responses

• 200 {} once the trace is stored, including for Test Connection's empty payload.
• 401 with a wrong or missing key.
• Other refusals are OTLP's, with Retry-After.

Each LLM call counts as one event.

Next steps

• Boards: build the spend board.
• Alerts: get told when llm_call errors spike.