@mupt-ai/dari-gateway 0.3.0 → 0.3.1

package/dist/providers/provider_reported_cost.js

@@ -2,7 +2,8 @@ import { isRecord } from "@mupt-ai/dari-router/policy-engine";
 import { gatewayFor } from "../models/gateways.js";
 import { providerFailureFromCaught } from "../runtime/provider_stream_outcome.js";
 // Gateways bill per call and say what they billed: OpenRouter as `usage.cost`
-// on the final chunk, Vercel AI Gateway as `providerMetadata.gateway.cost`.
+// on the final chunk (plus the provider's charge on a BYOK call), Vercel AI
+// Gateway as `providerMetadata.gateway.cost`.
 // Pi prices usage from model rates and drops both, so the amount is read off
 // the raw response bytes and replaces the priced total.
 export function reportsProviderCost(model) {
@@ -63,7 +64,21 @@ export function reportedCost(body) {
     const gateway = isRecord(body.providerMetadata) && isRecord(body.providerMetadata.gateway)
         ? body.providerMetadata.gateway
         : undefined;
-    return usd(isRecord(body.usage) ? body.usage.cost : undefined) ?? usd(gateway?.cost);
+    return openRouterCost(body.usage) ?? usd(gateway?.cost);
+}
+// On a BYOK call OpenRouter bills the provider's charge to the org's own key:
+// `usage.cost` is only its fee (often 0), and the charge is in
+// `cost_details.upstream_inference_cost`.
+function openRouterCost(usage) {
+    if (!isRecord(usage))
+        return undefined;
+    const cost = usd(usage.cost);
+    if (cost === undefined || usage.is_byok !== true)
+        return cost;
+    const upstream = isRecord(usage.cost_details)
+        ? usd(usage.cost_details.upstream_inference_cost)
+        : undefined;
+    return cost + (upstream ?? 0);
 }
 function usd(value) {
     const amount = typeof value === "string" && value.trim() !== "" ? Number(value) : value;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mupt-ai/dari-gateway",
-  "version": "0.3.0",
+  "version": "0.3.1",
   "description": "Self-hostable LLM gateway for coding agents and APIs: one endpoint for Anthropic Messages, OpenAI Chat Completions, and OpenAI Responses, with subscription pooling, failover, and prompt-cache-aware routing.",
   "license": "Apache-2.0",
   "repository": {

package/src/providers/provider_reported_cost.ts

@@ -5,7 +5,8 @@ import { providerFailureFromCaught } from "../runtime/provider_stream_outcome.js
 import type { ProviderFetch } from "./response_capture.js";
 
 // Gateways bill per call and say what they billed: OpenRouter as `usage.cost`
-// on the final chunk, Vercel AI Gateway as `providerMetadata.gateway.cost`.
+// on the final chunk (plus the provider's charge on a BYOK call), Vercel AI
+// Gateway as `providerMetadata.gateway.cost`.
 // Pi prices usage from model rates and drops both, so the amount is read off
 // the raw response bytes and replaces the priced total.
 
@@ -84,7 +85,20 @@ export function reportedCost(body: unknown): number | undefined {
   const gateway = isRecord(body.providerMetadata) && isRecord(body.providerMetadata.gateway)
     ? body.providerMetadata.gateway
     : undefined;
-  return usd(isRecord(body.usage) ? body.usage.cost : undefined) ?? usd(gateway?.cost);
+  return openRouterCost(body.usage) ?? usd(gateway?.cost);
+}
+
+// On a BYOK call OpenRouter bills the provider's charge to the org's own key:
+// `usage.cost` is only its fee (often 0), and the charge is in
+// `cost_details.upstream_inference_cost`.
+function openRouterCost(usage: unknown): number | undefined {
+  if (!isRecord(usage)) return undefined;
+  const cost = usd(usage.cost);
+  if (cost === undefined || usage.is_byok !== true) return cost;
+  const upstream = isRecord(usage.cost_details)
+    ? usd(usage.cost_details.upstream_inference_cost)
+    : undefined;
+  return cost + (upstream ?? 0);
 }
 
 function usd(value: unknown): number | undefined {