What does the gen_ai.usage.reasoning.output_tokens attribute mean in Rius?

In Rius, gen_ai.usage.reasoning.output_tokens is an integer on a model-call span that holds the reasoning tokens in that call. Those tokens are a subset of the span’s output token count, not an extra amount on top of it. You set it next to gen_ai.usage.output_tokens, and Rius stores it under that key.

Rius computes cost on its own from the request model and the token counts, so you never send a cost value. The Span attribute reference lists the key with the other usage counts.

How it works

• What the key holds. A whole number of reasoning tokens for one model call, such as the tokens a reasoning model spends thinking before it answers.
• A subset, not an addition. gen_ai.usage.output_tokens already includes the reasoning tokens. If a call has 900 output tokens and 600 of them are reasoning, you send 900 and 600. Do not add 600 to the output count.
• Which spans carry it. The span for a model call, next to gen_ai.request.model, gen_ai.usage.input_tokens and gen_ai.usage.output_tokens. The cache counts, gen_ai.usage.cache_read.input_tokens and gen_ai.usage.cache_write.input_tokens, are subsets of the input count in the same way. A span that is not a model call has no tokens to report.
• Cost. Rius prices a span from gen_ai.request.model and the input, output, cache read and cache write counts, each at its own rate. Reasoning tokens are already inside the output count, so you never send a cost attribute. A span with no model, a model the price catalog does not know, or none of the counts gets no cost. Costs are list-price estimates in US dollars.
• Older names. Pydantic AI reports usage details under gen_ai.usage.details.reasoning_tokens, thinking_tokens or thoughts_tokens, and OpenInference reports reasoning as a detail of llm.token_count.*. Rius moves those values onto gen_ai.usage.reasoning.output_tokens when it ingests the span and removes the source keys. A native gen_ai.* key always wins, because normalization fills gaps and never overwrites.
• Searching. Filter on gen_ai.usage.reasoning.output_tokens, the key the value is stored under. A filter on the Pydantic AI source name matches nothing.
• Who defines it. It is an OpenTelemetry GenAI convention key, listed in the registry.

Set it up

1. Send the reasoning count only when your provider or instrumentation reports it. A span with no reasoning count is still a valid model-call span, and Rius reads any usage count a span leaves out as 0 for cost.
2. Put the reasoning count on the same span as the request model and the output count, and keep it at or below the output count. A call with 1,200 prompt tokens, 900 output tokens and 600 reasoning tokens carries these attributes. The values are examples, so use your own model name and counts:

{
  "gen_ai.request.model": "your-model-name",
  "gen_ai.usage.input_tokens": 1200,
  "gen_ai.usage.output_tokens": 900,
  "gen_ai.usage.reasoning.output_tokens": 600
}

3. If you emit spans without the Rius SDK, the Span attribute reference shows where to send them (https://ingest.eu.console.rius-glassflow.com/v1/traces, over OTLP/HTTP with an API key as a Bearer header) and which other attributes to set. For the full tracing setup, see Trace LLM calls with OpenTelemetry.
4. If you use Pydantic AI, send nothing extra. Rius normalizes its gen_ai.usage.details.reasoning_tokens value at ingest.

What you get

• A reasoning token count stored on the model-call span, under one key, whenever the sender supplies a reasoning count in a form Rius reads.
• Output and reasoning counts that stay consistent, because reasoning is a subset of output and is never double counted in the output total.
• Cost from the model and the input, output and cache counts. The reasoning tokens are already part of the output count, so nothing extra is needed.
• Reasoning tokens as a measure in custom alert rules, next to input and output tokens. Tokens sit on model-call spans, so filter such a rule by Model, or by Span kind LLM, to measure them.
• One search key. Pydantic AI usage details are moved to gen_ai.usage.reasoning.output_tokens, so a single filter covers them.

When something else fits

• If your provider does not report reasoning tokens, there is nothing to set. The output count still drives cost.
• If you only need the total completion size, send gen_ai.usage.output_tokens and skip the reasoning key.
• If you are sending spans from another convention, such as OpenInference or OpenLLMetry, the Span attribute reference shows how each is normalized to the keys Rius searches.

Related