@juspay/neurolink 10.6.4 → 10.7.0

package/dist/utils/pricing.js

@@ -653,15 +653,27 @@ function findRates(provider, model) {
     if (!providerPricing) {
         return undefined;
     }
+    // Bedrock model IDs carry region/vendor prefixes ("us.anthropic.claude-…",
+    // "global.anthropic.claude-…", "anthropic.claude-…") or arrive as full
+    // inference-profile ARNs ("arn:aws:bedrock:…:inference-profile/us.anthropic.
+    // claude-…" — the form this repo's own setup guides recommend). None of
+    // those match the bare pricing keys ("claude-…"), so every Bedrock call
+    // priced to $0. Strip everything up to and including the vendor prefix; the
+    // trailing "-v1:0" version suffix is absorbed by the longest-prefix match
+    // below. Non-Anthropic Bedrock models (meta./amazon./mistral.) don't match
+    // and fall through unchanged, exactly as before.
+    const modelKey = stripped === "bedrock" || stripped === "amazonbedrock"
+        ? model.replace(/^.*\banthropic\./, "")
+        : model;
     // Exact match
-    if (providerPricing[model]) {
-        return providerPricing[model];
+    if (providerPricing[modelKey]) {
+        return providerPricing[modelKey];
     }
     // Longest-prefix match (skip the synthetic "_default" sentinel below)
     const sortedKeys = Object.keys(providerPricing)
         .filter((k) => k !== "_default")
         .sort((a, b) => b.length - a.length);
-    const key = sortedKeys.find((k) => model.startsWith(k));
+    const key = sortedKeys.find((k) => modelKey.startsWith(k));
     if (key) {
         return providerPricing[key];
     }

package/dist/utils/tokenUtils.js

@@ -44,10 +44,13 @@ export function extractOutputTokens(usage) {
  * Falls back to input + output if total is not provided
  */
 export function extractTotalTokens(usage, input, output) {
-    if (typeof usage.total === "number") {
+    // A literal 0 alongside non-zero components means the provider omitted the
+    // figure (several gateways/parsers emit total: 0) — fall through to the
+    // computed sum instead of letting the zero win.
+    if (typeof usage.total === "number" && usage.total > 0) {
         return usage.total;
     }
-    if (typeof usage.totalTokens === "number") {
+    if (typeof usage.totalTokens === "number" && usage.totalTokens > 0) {
         return usage.totalTokens;
     }
     return input + output;
@@ -185,7 +188,6 @@ export function extractTokenUsage(result, options = {}) {
     // Extract base token counts
     let input = extractInputTokens(usage);
     const output = extractOutputTokens(usage);
-    const total = extractTotalTokens(usage, input, output);
     // Extract optional token fields
     const reasoning = extractReasoningTokens(usage);
     const cacheCreationTokens = extractCacheCreationTokens(usage);
@@ -204,6 +206,39 @@ export function extractTokenUsage(result, options = {}) {
             input = Math.max(0, input - overlappingCached);
         }
     }
+    // ai@6-shape rebase: when the cache values came ONLY from the ai@6
+    // normalized shape (cachedInputTokens / inputTokenDetails.*), the flat
+    // input is cache-INCLUSIVE — asLanguageModelUsage reports
+    // inputTokens.total = noCache + cacheRead + cacheWrite — so `input` must
+    // be rebased onto the uncached portion or calculateCost bills the cached
+    // tokens twice (full input rate + cacheRead/cacheCreation rate). The
+    // native fields excluded below (cacheReadInputTokens / cacheReadTokens /
+    // cacheCreationInputTokens / cacheCreationTokens) follow the
+    // non-overlapping convention where input is already the uncached
+    // remainder, and the overlapping snake_case shape was already rebased
+    // above — neither must be rebased again.
+    if ((cacheReadTokens !== undefined || cacheCreationTokens !== undefined) &&
+        usage.cacheReadInputTokens === undefined &&
+        usage.cacheReadTokens === undefined &&
+        usage.cacheCreationInputTokens === undefined &&
+        usage.cacheCreationTokens === undefined &&
+        usage.prompt_tokens_details?.cached_tokens === undefined &&
+        (usage.cachedInputTokens !== undefined ||
+            usage.inputTokenDetails !== undefined)) {
+        const noCacheTokens = usage.inputTokenDetails?.noCacheTokens;
+        if (typeof noCacheTokens === "number") {
+            input = noCacheTokens;
+        }
+        else {
+            input = Math.max(0, input -
+                (cacheReadTokens ?? 0) -
+                (usage.inputTokenDetails?.cacheWriteTokens ?? 0));
+        }
+    }
+    // Total: prefer the provider-reported figure; the fallback conserves every
+    // billed component — cache tokens are additive on top of the (post-rebase)
+    // uncached input. Reasoning is already inside `output`, so it is NOT added.
+    const total = extractTotalTokens(usage, input + (cacheReadTokens ?? 0) + (cacheCreationTokens ?? 0), output);
     // Calculate cache savings if enabled
     const cacheSavingsPercent = calculateCacheSavings
         ? calculateCacheSavingsPercent(cacheReadTokens, input)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "10.6.4",
+  "version": "10.7.0",
   "packageManager": "pnpm@10.15.1",
   "description": "TypeScript AI SDK with 24+ LLM providers behind one consistent API. MCP-native (connect any MCP server), voice TTS/STT/realtime, RAG, agents, memory, context compaction. OpenAI · Anthropic · Gemini · Bedrock · Azure · Ollama · DeepSeek · NVIDIA NIM and more.",
   "author": {
@@ -104,6 +104,7 @@
     "test:google-native": "npx tsx test/continuous-test-suite-google-native.ts",
     "test:gemini-abort": "npx tsx test/continuous-test-suite-gemini-abort.ts",
     "test:anthropic-cap": "npx tsx test/continuous-test-suite-anthropic-cap.ts",
+    "test:token-usage": "npx tsx test/continuous-test-suite-token-usage.ts",
     "test:tts": "npx tsx test/continuous-test-suite-tts.ts",
     "test:voice": "npx tsx test/continuous-test-suite-voice.ts",
     "test:voice-server": "npx tsx test/continuous-test-suite-voice-server.ts",