Turn LLM usage records into OpenTelemetry GenAI spans with computed per-call cost. Feed it what your app already logs — provider, model, token counts — and get back OTLP/JSON spans (ready to ship to any OpenTelemetry endpoint) plus a cost summary.

Part of the Kinetic Gain GenAI observability lane.

Most teams log token counts somewhere but can't answer "what did last night's agent run cost, by model?" without a spreadsheet. The OpenTelemetry GenAI semantic conventions standardize how to describe an LLM call as a span (gen_ai.request.model, gen_ai.usage.input_tokens, …) but say nothing about cost. This exporter bridges that gap: it emits spec-correct GenAI spans and attaches cost as a documented extension attribute, so cost shows up next to latency and tokens in the same trace backend you already run.

It is offline and deterministic — it transforms records, it does not call any LLM or network endpoint.

npm install -g llm-cost-span-exporter   # CLI
npm install llm-cost-span-exporter      # library

Requires Node ≥ 20.

# OTLP spans to stdout, cost summary to stderr
llm-cost-export usage.jsonl --summary

# write OTLP to a file, using your own price table
llm-cost-export usage.jsonl --prices my-prices.json --out spans.json

Input is a JSON array, a single JSON object, or JSONL of usage records:

{ "provider": "anthropic", "model": "claude-opus-4-7", "operation": "chat", "inputTokens": 12000, "outputTokens": 3400, "durationMs": 4200 }

import { exportSpans } from "llm-cost-span-exporter";

const { otlp, summary } = exportSpans(records);
console.log(summary.totalCost, summary.byModel);
// POST otlp to your OTLP/HTTP traces endpoint

Standard OTel GenAI (per the spec):

Cost extension attributes (non-standard — OTel has no official cost attribute):

gen_ai.usage.cost, gen_ai.usage.input_cost, gen_ai.usage.output_cost, gen_ai.usage.cost_currency, and gen_ai.usage.cost_unpriced (set when the model isn't in the price table).

A small indicative price table ships built in (USD per 1M tokens, point-in-time). Prices change constantly — override with --prices file.json:

{ "claude-opus-4": { "inputPerMTok": 15, "outputPerMTok": 75 } }

Model ids resolve by exact match, then by the longest price-table key that is a substring (so gpt-4o-2024-11-20 prices via gpt-4o). Unknown models are reported as unpriced with zero cost rather than silently mispriced.

AGPL-3.0-or-later — see LICENSE.