ai-lcr 0.6.5 → 0.7.0

package/CHANGELOG.md

@@ -4,6 +4,44 @@ All notable changes to `ai-lcr` are documented here. The format follows
 [Keep a Changelog](https://keepachangelog.com/), and the project adheres to
 [Semantic Versioning](https://semver.org/).
 
+## [0.7.0] — 2026-06-20
+
+The text router now records the **provider-reported actual cost** when a provider
+returns one, instead of always estimating from the price table. The table becomes
+the routing input and the drift baseline (`estCostUsd`); the recorded `costUsd` is
+the real bill wherever the provider gives it.
+
+### Why
+
+A static price table can only encode one price per model, but an aggregator
+(OpenRouter) routes a single model across many sub-providers whose prices differ
+several-fold, picking one per call — so `tokens × table` is structurally unable to
+match the bill for multi-provider models (measured: `deepseek-v4-pro` reconciled at
+~57% of the real cost, while single-provider models like Gemini/Claude/GPT matched
+at 100%). The provider's own number already accounts for which sub-provider served,
+every token kind (cache read/write, reasoning), and fees — none of which a flat
+table can track.
+
+### Added
+
+- **`costUsd` prefers the provider-reported actual cost** (text path). Read from
+  OpenRouter's `providerMetadata.openrouter.usage` —
+  `costDetails.upstreamInferenceCost` (the real upstream / BYOK model spend) when
+  present, otherwise `cost` (the credit charge) — and from an OpenAI-compatible
+  provider's `estimated_cost` on the raw usage body. Requires the caller to enable
+  usage accounting on the provider (e.g. OpenRouter `usage: { include: true }`);
+  without it, behavior is unchanged.
+- **`estCostUsd` is now set on text records** (previously media-only) — the
+  price-table prediction for the same usage. `costUsd − estCostUsd` is the
+  price-table drift signal, so a dashboard's drift panel now works for text too.
+
+### Changed
+
+- When no provider cost is reported, `costUsd` still equals the price-table
+  estimate (and `estCostUsd` equals it, so no drift is flagged) — a pure fallback,
+  fully backward-compatible. The streaming path reads the reported cost from the
+  `finish` chunk's `providerMetadata`.
+
 ## [0.6.5] — 2026-06-16
 
 Bundled price table now covers the open-weights labs, not just the Western

package/dist/index.cjs

@@ -341,6 +341,20 @@ function cacheSavingForUsage(cost, inputTokens, cacheReadTokens) {
   const cached = Math.min(Math.max(cacheReadTokens, 0), inputTokens);
   return cached / 1e6 * (cost.input - cost.cacheRead);
 }
+function reportedCost(providerMetadata, usage) {
+  const orUsage = providerMetadata?.openrouter?.usage;
+  if (orUsage) {
+    const upstream = orUsage.costDetails?.upstreamInferenceCost;
+    if (typeof upstream === "number" && upstream > 0) return upstream;
+    if (typeof orUsage.cost === "number") return orUsage.cost;
+  }
+  const raw = usage?.raw;
+  if (raw) {
+    const est = raw["estimated_cost"] ?? raw["cost"];
+    if (typeof est === "number") return est;
+  }
+  return void 0;
+}
 function requestIdFrom(options) {
   const raw = options.providerOptions?.lcr?.requestId;
   return typeof raw === "string" && raw.length > 0 ? raw : void 0;
@@ -539,12 +553,13 @@ var LcrFallbackModel = class {
     return baseline;
   }
   /** Winner settled: record the attempt, fire `onCost` (compat) + `onCall`. */
-  finalizeOk(ctx, provider, attemptStart, usage, ttftMs) {
+  finalizeOk(ctx, provider, attemptStart, usage, ttftMs, providerMetadata) {
     ctx.attempts.push({ provider: provider.label, ok: true, latencyMs: Date.now() - attemptStart });
     const inputTokens = usage?.inputTokens?.total ?? 0;
     const outputTokens = usage?.outputTokens?.total ?? 0;
     const cacheReadTokens = usage?.inputTokens?.cacheRead ?? 0;
-    const costUsd = provider.cost ? costForUsage(provider.cost, inputTokens, outputTokens, cacheReadTokens) : 0;
+    const estCostUsd = provider.cost ? costForUsage(provider.cost, inputTokens, outputTokens, cacheReadTokens) : void 0;
+    const costUsd = reportedCost(providerMetadata, usage) ?? estCostUsd ?? 0;
     const cachedSavingUsd = provider.cost ? cacheSavingForUsage(provider.cost, inputTokens, cacheReadTokens) : 0;
     const usageMissing = inputTokens === 0 && outputTokens === 0;
     const emptyCompletion = inputTokens > 0 && outputTokens === 0;
@@ -579,6 +594,7 @@ var LcrFallbackModel = class {
       outputTokens,
       ...cacheReadTokens > 0 ? { cachedInputTokens: cacheReadTokens } : {},
       costUsd,
+      ...estCostUsd !== void 0 ? { estCostUsd } : {},
       ...baselineUsd !== void 0 ? { baselineUsd, baselineKind: "last-leg" } : {},
       ...cachedSavingUsd > 0 ? { cachedSavingUsd } : {},
       ...ctx.requestId ? { requestId: ctx.requestId } : {},
@@ -635,7 +651,7 @@ var LcrFallbackModel = class {
         }
         this.recordProviderSuccess(idx);
         this.settleSticky(idx);
-        this.finalizeOk(ctx, provider, attemptStart, result.usage);
+        this.finalizeOk(ctx, provider, attemptStart, result.usage, void 0, result.providerMetadata);
         if (cache && cacheKey !== void 0 && ctx.settled?.cacheable) {
           this.storeCache(cacheKey, { kind: "generate", result, meta: ctx.settled.meta });
         }
@@ -767,6 +783,7 @@ var LcrFallbackModel = class {
     const servingIdx = idx;
     const servingPos = p;
     let usage;
+    let finishProviderMetadata;
     let contentStreamed = false;
     let ttftMs;
     const stream = new ReadableStream({
@@ -783,6 +800,7 @@ var LcrFallbackModel = class {
             if (done) break;
             if (value.type === "finish") {
               usage = value.usage;
+              finishProviderMetadata = value.providerMetadata;
               const out = value.usage?.outputTokens?.total ?? 0;
               const inp = value.usage?.inputTokens?.total ?? 0;
               if (inp > 0 && out === 0 && !contentStreamed && servingPos + 1 < n) {
@@ -797,7 +815,7 @@ var LcrFallbackModel = class {
           }
           self.recordProviderSuccess(servingIdx);
           self.settleSticky(servingIdx);
-          self.finalizeOk(ctx, servingProvider, servingAttemptStart, usage, ttftMs);
+          self.finalizeOk(ctx, servingProvider, servingAttemptStart, usage, ttftMs, finishProviderMetadata);
           controller.close();
         } catch (error) {
           self.emitError(error, servingProvider.label);

package/dist/index.js

@@ -287,6 +287,20 @@ function cacheSavingForUsage(cost, inputTokens, cacheReadTokens) {
   const cached = Math.min(Math.max(cacheReadTokens, 0), inputTokens);
   return cached / 1e6 * (cost.input - cost.cacheRead);
 }
+function reportedCost(providerMetadata, usage) {
+  const orUsage = providerMetadata?.openrouter?.usage;
+  if (orUsage) {
+    const upstream = orUsage.costDetails?.upstreamInferenceCost;
+    if (typeof upstream === "number" && upstream > 0) return upstream;
+    if (typeof orUsage.cost === "number") return orUsage.cost;
+  }
+  const raw = usage?.raw;
+  if (raw) {
+    const est = raw["estimated_cost"] ?? raw["cost"];
+    if (typeof est === "number") return est;
+  }
+  return void 0;
+}
 function requestIdFrom(options) {
   const raw = options.providerOptions?.lcr?.requestId;
   return typeof raw === "string" && raw.length > 0 ? raw : void 0;
@@ -485,12 +499,13 @@ var LcrFallbackModel = class {
     return baseline;
   }
   /** Winner settled: record the attempt, fire `onCost` (compat) + `onCall`. */
-  finalizeOk(ctx, provider, attemptStart, usage, ttftMs) {
+  finalizeOk(ctx, provider, attemptStart, usage, ttftMs, providerMetadata) {
     ctx.attempts.push({ provider: provider.label, ok: true, latencyMs: Date.now() - attemptStart });
     const inputTokens = usage?.inputTokens?.total ?? 0;
     const outputTokens = usage?.outputTokens?.total ?? 0;
     const cacheReadTokens = usage?.inputTokens?.cacheRead ?? 0;
-    const costUsd = provider.cost ? costForUsage(provider.cost, inputTokens, outputTokens, cacheReadTokens) : 0;
+    const estCostUsd = provider.cost ? costForUsage(provider.cost, inputTokens, outputTokens, cacheReadTokens) : void 0;
+    const costUsd = reportedCost(providerMetadata, usage) ?? estCostUsd ?? 0;
     const cachedSavingUsd = provider.cost ? cacheSavingForUsage(provider.cost, inputTokens, cacheReadTokens) : 0;
     const usageMissing = inputTokens === 0 && outputTokens === 0;
     const emptyCompletion = inputTokens > 0 && outputTokens === 0;
@@ -525,6 +540,7 @@ var LcrFallbackModel = class {
       outputTokens,
       ...cacheReadTokens > 0 ? { cachedInputTokens: cacheReadTokens } : {},
       costUsd,
+      ...estCostUsd !== void 0 ? { estCostUsd } : {},
       ...baselineUsd !== void 0 ? { baselineUsd, baselineKind: "last-leg" } : {},
       ...cachedSavingUsd > 0 ? { cachedSavingUsd } : {},
       ...ctx.requestId ? { requestId: ctx.requestId } : {},
@@ -581,7 +597,7 @@ var LcrFallbackModel = class {
         }
         this.recordProviderSuccess(idx);
         this.settleSticky(idx);
-        this.finalizeOk(ctx, provider, attemptStart, result.usage);
+        this.finalizeOk(ctx, provider, attemptStart, result.usage, void 0, result.providerMetadata);
         if (cache && cacheKey !== void 0 && ctx.settled?.cacheable) {
           this.storeCache(cacheKey, { kind: "generate", result, meta: ctx.settled.meta });
         }
@@ -713,6 +729,7 @@ var LcrFallbackModel = class {
     const servingIdx = idx;
     const servingPos = p;
     let usage;
+    let finishProviderMetadata;
     let contentStreamed = false;
     let ttftMs;
     const stream = new ReadableStream({
@@ -729,6 +746,7 @@ var LcrFallbackModel = class {
             if (done) break;
             if (value.type === "finish") {
               usage = value.usage;
+              finishProviderMetadata = value.providerMetadata;
               const out = value.usage?.outputTokens?.total ?? 0;
               const inp = value.usage?.inputTokens?.total ?? 0;
               if (inp > 0 && out === 0 && !contentStreamed && servingPos + 1 < n) {
@@ -743,7 +761,7 @@ var LcrFallbackModel = class {
           }
           self.recordProviderSuccess(servingIdx);
           self.settleSticky(servingIdx);
-          self.finalizeOk(ctx, servingProvider, servingAttemptStart, usage, ttftMs);
+          self.finalizeOk(ctx, servingProvider, servingAttemptStart, usage, ttftMs, finishProviderMetadata);
           controller.close();
         } catch (error) {
           self.emitError(error, servingProvider.label);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ai-lcr",
-  "version": "0.6.5",
+  "version": "0.7.0",
   "description": "Least Cost Routing for LLMs — route every model call to the cheapest available provider, fall back automatically, and track real cost. Built for the Vercel AI SDK.",
   "keywords": [
     "ai",