@juspay/neurolink 10.6.4 → 10.6.5

package/dist/core/analytics.js

@@ -98,8 +98,14 @@ function estimateCost(provider, model, tokens) {
         if (!costInfo) {
             return undefined;
         }
-        // Calculate cost using the configuration system (per-1K-token rates)
-        const inputCost = (tokens.input / 1000) * costInfo.input;
+        // Calculate cost using the configuration system (per-1K-token rates).
+        // costInfo has no cache tiers, so cache tokens are billed at the input
+        // rate — the pre-split total a cache-blind provider would report, never $0.
+        const inputCost = ((tokens.input +
+            (tokens.cacheReadTokens ?? 0) +
+            (tokens.cacheCreationTokens ?? 0)) /
+            1000) *
+            costInfo.input;
         const outputCost = (tokens.output / 1000) * costInfo.output;
         return Math.round((inputCost + outputCost) * 1_000_000) / 1_000_000; // Round to 6 decimal places
     }

package/dist/core/baseProvider.js

@@ -27,6 +27,7 @@ import { TelemetryHandler } from "./modules/TelemetryHandler.js";
 import { ToolsManager } from "./modules/ToolsManager.js";
 import { Utilities } from "./modules/Utilities.js";
 import { generateText } from "../utils/generation.js";
+import { extractTokenUsage } from "../utils/tokenUtils.js";
 /**
  * Read the consumer-facing lifecycle callbacks buried inside a request's
  * middleware blob. The parameter is `unknown` on purpose: request options
@@ -1079,12 +1080,7 @@ export class BaseProvider {
                     experimental_telemetry: this.telemetryHandler?.getTelemetryConfig(options, "generate"),
                 });
                 formattedContent = formattedResult.text;
-                usage = {
-                    input: formattedResult.usage?.inputTokens || 0,
-                    output: formattedResult.usage?.outputTokens || 0,
-                    total: (formattedResult.usage?.inputTokens || 0) +
-                        (formattedResult.usage?.outputTokens || 0),
-                };
+                usage = extractTokenUsage(formattedResult.totalUsage ?? formattedResult.usage);
                 logger.debug("[VideoAnalysis] Claude formatting complete", {
                     formattedLength: formattedContent.length,
                     usage,
@@ -1127,13 +1123,19 @@ export class BaseProvider {
         this.analyzeAIResponse(generateResult);
         this.logGenerationComplete(generateResult);
         const responseTime = Date.now() - startTime;
-        await this.recordPerformanceMetrics(generateResult.usage, responseTime);
         const { toolsUsed, toolExecutions } = this.extractToolInformation(generateResult);
         // Prefer the per-call recorder's real records (params/result/timing per
         // execution); fall back to a conversion of the step-extraction entries
         // for tools the recorder could not wrap (provider-executed tools).
         const toolExecutionRecords = resolveToolExecutionRecords(options, toolExecutions);
         let enhancedResult = this.formatEnhancedResult(generateResult, tools, toolsUsed, toolExecutionRecords, options);
+        // Recorded AFTER formatEnhancedResult so telemetry sees the same usage
+        // the caller gets: the cross-step aggregate (totalUsage, not last-step
+        // usage) WITH the providerMetadata cache merge applied — otherwise
+        // providers whose cache data lives only in providerMetadata would have
+        // their cache tokens billed at the full input rate in OTEL metrics,
+        // diverging from analytics.cost.
+        await this.recordPerformanceMetrics(enhancedResult.usage, responseTime);
         enhancedResult = await this.synthesizeAIResponseIfNeeded(enhancedResult, options);
         const finalResult = await this.enhanceResult(enhancedResult, options, startTime);
         return finalResult;

package/dist/core/modules/GenerationHandler.d.ts

@@ -68,6 +68,14 @@ export declare class GenerationHandler {
      * The AI SDK's LanguageModelUsage only has inputTokens/outputTokens.
      * Cache metrics are surfaced via providerMetadata by provider-specific SDK adapters.
      */
+    /**
+     * Set gen_ai usage attributes + cache-aware cost on the span from the
+     * CROSS-STEP aggregate (result.totalUsage). result.usage is the LAST step
+     * only — using it undercounted every multi-step tool loop, and pricing the
+     * raw cache-inclusive inputTokens without the cache fields billed cache
+     * reads at the full input rate.
+     */
+    private setUsageSpanAttributes;
     private extractCacheMetricsFromProviderMetadata;
     /**
      * Log generation completion information

package/dist/core/modules/GenerationHandler.js

@@ -475,18 +475,7 @@ export class GenerationHandler {
                     });
                 }
                 // Set token usage and completion attributes on span
-                if (result.usage) {
-                    span.setAttribute("gen_ai.usage.input_tokens", result.usage.inputTokens || 0);
-                    span.setAttribute("gen_ai.usage.output_tokens", result.usage.outputTokens || 0);
-                    // Cost on span so users can query "what did this trace cost?"
-                    const cost = calculateCost(this.providerName, this.modelName, {
-                        input: result.usage.inputTokens || 0,
-                        output: result.usage.outputTokens || 0,
-                        total: (result.usage.inputTokens || 0) +
-                            (result.usage.outputTokens || 0),
-                    });
-                    span.setAttribute("neurolink.cost", cost ?? 0);
-                }
+                this.setUsageSpanAttributes(span, result);
                 if (result.finishReason) {
                     span.setAttribute("gen_ai.response.finish_reason", result.finishReason);
                 }
@@ -559,17 +548,7 @@ export class GenerationHandler {
                         toolCallsTotal: result.toolCalls?.length || 0,
                         responseChars: result.text?.length || 0,
                     });
-                    if (result.usage) {
-                        span.setAttribute("gen_ai.usage.input_tokens", result.usage.inputTokens || 0);
-                        span.setAttribute("gen_ai.usage.output_tokens", result.usage.outputTokens || 0);
-                        const fallbackCost = calculateCost(this.providerName, this.modelName, {
-                            input: result.usage.inputTokens || 0,
-                            output: result.usage.outputTokens || 0,
-                            total: (result.usage.inputTokens || 0) +
-                                (result.usage.outputTokens || 0),
-                        });
-                        span.setAttribute("neurolink.cost", fallbackCost ?? 0);
-                    }
+                    this.setUsageSpanAttributes(span, result);
                     if (result.finishReason) {
                         span.setAttribute("gen_ai.response.finish_reason", result.finishReason);
                     }
@@ -606,17 +585,7 @@ export class GenerationHandler {
                         "retry.strategy": "temperature_omitted",
                     });
                     span.setAttribute("retry.count", 1);
-                    if (result.usage) {
-                        span.setAttribute("gen_ai.usage.input_tokens", result.usage.inputTokens || 0);
-                        span.setAttribute("gen_ai.usage.output_tokens", result.usage.outputTokens || 0);
-                        const noTempCost = calculateCost(this.providerName, this.modelName, {
-                            input: result.usage.inputTokens || 0,
-                            output: result.usage.outputTokens || 0,
-                            total: (result.usage.inputTokens || 0) +
-                                (result.usage.outputTokens || 0),
-                        });
-                        span.setAttribute("neurolink.cost", noTempCost ?? 0);
-                    }
+                    this.setUsageSpanAttributes(span, result);
                     if (result.finishReason) {
                         span.setAttribute("gen_ai.response.finish_reason", result.finishReason);
                     }
@@ -640,6 +609,26 @@ export class GenerationHandler {
      * The AI SDK's LanguageModelUsage only has inputTokens/outputTokens.
      * Cache metrics are surfaced via providerMetadata by provider-specific SDK adapters.
      */
+    /**
+     * Set gen_ai usage attributes + cache-aware cost on the span from the
+     * CROSS-STEP aggregate (result.totalUsage). result.usage is the LAST step
+     * only — using it undercounted every multi-step tool loop, and pricing the
+     * raw cache-inclusive inputTokens without the cache fields billed cache
+     * reads at the full input rate.
+     */
+    setUsageSpanAttributes(span, result) {
+        const aggregate = result.totalUsage ?? result.usage;
+        if (!aggregate) {
+            return;
+        }
+        span.setAttribute("gen_ai.usage.input_tokens", aggregate.inputTokens || 0);
+        span.setAttribute("gen_ai.usage.output_tokens", aggregate.outputTokens || 0);
+        // Cost on span so users can query "what did this trace cost?" —
+        // extractTokenUsage rebases input onto the uncached remainder and
+        // surfaces the cache fields so calculateCost prices each tier.
+        const cost = calculateCost(this.providerName, this.modelName, extractTokenUsage(aggregate));
+        span.setAttribute("neurolink.cost", cost ?? 0);
+    }
     extractCacheMetricsFromProviderMetadata(generateResult) {
         const providerMeta = generateResult.providerMetadata;
         if (!providerMeta) {
@@ -841,10 +830,11 @@ export class GenerationHandler {
             jsonTruncated = true;
             logger.warn("[GenerationHandler] Structured output truncated by token cap (finishReason=length); increase maxTokens", { provider: this.providerName, model: this.modelName });
         }
-        // Extract usage with support for different formats and reasoning tokens
-        // Note: The AI SDK bundles thinking tokens into promptTokens for Google models.
-        // Separate reasoningTokens tracking will work when/if the AI SDK adds support.
-        const usage = extractTokenUsage(generateResult.usage);
+        // Extract usage with support for different formats and reasoning tokens.
+        // totalUsage is the CROSS-STEP aggregate; generateResult.usage is the
+        // LAST step only, which silently dropped every prior step of a
+        // multi-step tool loop.
+        const usage = extractTokenUsage(generateResult.totalUsage ?? generateResult.usage);
         // Merge cache metrics from providerMetadata if not already present in usage
         // The AI SDK's LanguageModelUsage doesn't include cache tokens; they come from
         // provider-specific metadata (e.g. Anthropic's providerMetadata.anthropic)

package/dist/core/modules/TelemetryHandler.d.ts

@@ -14,7 +14,7 @@
  * @module core/modules/TelemetryHandler
  */
 import type { NeuroLink } from "../../neurolink.js";
-import type { StreamOptions, AIProviderName, AnalyticsData, EnhancedGenerateResult, EvaluationData, TextGenerationOptions } from "../../types/index.js";
+import type { StreamOptions, AIProviderName, AnalyticsData, EnhancedGenerateResult, EvaluationData, RawUsageObject, TextGenerationOptions, TokenUsage } from "../../types/index.js";
 /**
  * TelemetryHandler class - Handles analytics and telemetry for AI providers
  */
@@ -34,10 +34,7 @@ export declare class TelemetryHandler {
     /**
      * Record performance metrics for a generation
      */
-    recordPerformanceMetrics(usage: {
-        inputTokens: number | undefined;
-        outputTokens: number | undefined;
-    } | undefined, responseTime: number): Promise<void>;
+    recordPerformanceMetrics(usage: RawUsageObject | undefined, responseTime: number): Promise<void>;
     /**
      * Calculate actual cost based on token usage and provider configuration.
      *
@@ -50,10 +47,7 @@ export declare class TelemetryHandler {
      * causing a ~1,780x under-estimate when the actual model was Claude Sonnet
      * on Vertex AI ($0.000060 vs $0.106895 for the same request).
      */
-    calculateActualCost(usage: {
-        inputTokens?: number | undefined;
-        outputTokens?: number | undefined;
-    }): Promise<number>;
+    calculateActualCost(usage: TokenUsage): Promise<number>;
     /**
      * Create telemetry configuration for Vercel AI SDK experimental_telemetry
      * This enables automatic OpenTelemetry tracing when telemetry is enabled

package/dist/core/modules/TelemetryHandler.js

@@ -14,6 +14,7 @@
  * @module core/modules/TelemetryHandler
  */
 import { nanoid } from "nanoid";
+import { extractTokenUsage } from "../../utils/tokenUtils.js";
 import { logger } from "../../utils/logger.js";
 import { recordProviderPerformanceFromMetrics } from "../evaluationProviders.js";
 import { modelConfig } from "../modelConfiguration.js";
@@ -75,8 +76,12 @@ export class TelemetryHandler {
      */
     async recordPerformanceMetrics(usage, responseTime) {
         try {
-            const totalTokens = (usage?.inputTokens || 0) + (usage?.outputTokens || 0);
-            const actualCost = await this.calculateActualCost(usage || { inputTokens: 0, outputTokens: 0 });
+            // Normalize first: rebases cache-inclusive ai@6 input onto the uncached
+            // remainder and surfaces cache tiers so the cost is priced per tier
+            // instead of billing cache reads at the full input rate.
+            const tokenUsage = extractTokenUsage(usage);
+            const totalTokens = tokenUsage.total;
+            const actualCost = await this.calculateActualCost(tokenUsage);
             recordProviderPerformanceFromMetrics(this.providerName, {
                 responseTime,
                 tokensGenerated: totalTokens,
@@ -109,25 +114,25 @@ export class TelemetryHandler {
      */
     async calculateActualCost(usage) {
         try {
-            const promptTokens = usage?.inputTokens || 0;
-            const completionTokens = usage?.outputTokens || 0;
             // Try the per-model pricing table first (includes correct rates for
             // Claude on Vertex, cache token rates, etc.)
             if (hasPricing(this.providerName, this.modelName)) {
-                return calculateCost(this.providerName, this.modelName, {
-                    input: promptTokens,
-                    output: completionTokens,
-                    total: promptTokens + completionTokens,
-                });
+                return calculateCost(this.providerName, this.modelName, usage);
             }
             // Fall back to provider-level default cost from configuration system
             const costInfo = modelConfig.getCostInfo(this.providerName, this.modelName);
             if (!costInfo) {
                 return 0; // No cost info available
             }
-            // Calculate cost per 1K tokens
-            const inputCost = (promptTokens / 1000) * costInfo.input;
-            const outputCost = (completionTokens / 1000) * costInfo.output;
+            // Calculate cost per 1K tokens. costInfo has no cache tiers, so cache
+            // tokens are billed at the input rate — the same total a provider
+            // without cache-aware splitting would have reported, never $0.
+            const inputCost = ((usage.input +
+                (usage.cacheReadTokens ?? 0) +
+                (usage.cacheCreationTokens ?? 0)) /
+                1000) *
+                costInfo.input;
+            const outputCost = (usage.output / 1000) * costInfo.output;
             return inputCost + outputCost;
         }
         catch (error) {

package/dist/lib/analytics/pricing.d.ts

@@ -21,4 +21,7 @@
  *   cross-provider search via openrouter alias)
  * @returns Total cost in USD
  */
-export declare function calculateAdvancedCost(model: string | undefined, inputTokens: number, outputTokens: number, provider?: string): number;
+export declare function calculateAdvancedCost(model: string | undefined, inputTokens: number, outputTokens: number, provider?: string, cacheTokens?: {
+    cacheReadTokens?: number;
+    cacheCreationTokens?: number;
+}): number;

package/dist/lib/analytics/pricing.js

@@ -22,17 +22,21 @@ import { calculateCost } from "../utils/pricing.js";
  *   cross-provider search via openrouter alias)
  * @returns Total cost in USD
  */
-export function calculateAdvancedCost(model, inputTokens, outputTokens, provider) {
+export function calculateAdvancedCost(model, inputTokens, outputTokens, provider, cacheTokens) {
     if (!model) {
         return 0;
     }
     // When provider is unknown, use the openrouter/litellm cross-provider
     // search path in findRates (PROVIDER_ALIASES → __cross_provider__).
     const resolvedProvider = provider && provider.length > 0 ? provider : "openrouter";
+    const cacheReadTokens = cacheTokens?.cacheReadTokens ?? 0;
+    const cacheCreationTokens = cacheTokens?.cacheCreationTokens ?? 0;
     return calculateCost(resolvedProvider, model, {
         input: inputTokens,
         output: outputTokens,
-        total: inputTokens + outputTokens,
+        total: inputTokens + cacheReadTokens + cacheCreationTokens + outputTokens,
+        ...(cacheReadTokens > 0 ? { cacheReadTokens } : {}),
+        ...(cacheCreationTokens > 0 ? { cacheCreationTokens } : {}),
     });
 }
 //# sourceMappingURL=pricing.js.map

package/dist/lib/core/analytics.js

@@ -98,8 +98,14 @@ function estimateCost(provider, model, tokens) {
         if (!costInfo) {
             return undefined;
         }
-        // Calculate cost using the configuration system (per-1K-token rates)
-        const inputCost = (tokens.input / 1000) * costInfo.input;
+        // Calculate cost using the configuration system (per-1K-token rates).
+        // costInfo has no cache tiers, so cache tokens are billed at the input
+        // rate — the pre-split total a cache-blind provider would report, never $0.
+        const inputCost = ((tokens.input +
+            (tokens.cacheReadTokens ?? 0) +
+            (tokens.cacheCreationTokens ?? 0)) /
+            1000) *
+            costInfo.input;
         const outputCost = (tokens.output / 1000) * costInfo.output;
         return Math.round((inputCost + outputCost) * 1_000_000) / 1_000_000; // Round to 6 decimal places
     }

package/dist/lib/core/baseProvider.js

@@ -27,6 +27,7 @@ import { TelemetryHandler } from "./modules/TelemetryHandler.js";
 import { ToolsManager } from "./modules/ToolsManager.js";
 import { Utilities } from "./modules/Utilities.js";
 import { generateText } from "../utils/generation.js";
+import { extractTokenUsage } from "../utils/tokenUtils.js";
 /**
  * Read the consumer-facing lifecycle callbacks buried inside a request's
  * middleware blob. The parameter is `unknown` on purpose: request options
@@ -1079,12 +1080,7 @@ export class BaseProvider {
                     experimental_telemetry: this.telemetryHandler?.getTelemetryConfig(options, "generate"),
                 });
                 formattedContent = formattedResult.text;
-                usage = {
-                    input: formattedResult.usage?.inputTokens || 0,
-                    output: formattedResult.usage?.outputTokens || 0,
-                    total: (formattedResult.usage?.inputTokens || 0) +
-                        (formattedResult.usage?.outputTokens || 0),
-                };
+                usage = extractTokenUsage(formattedResult.totalUsage ?? formattedResult.usage);
                 logger.debug("[VideoAnalysis] Claude formatting complete", {
                     formattedLength: formattedContent.length,
                     usage,
@@ -1127,13 +1123,19 @@ export class BaseProvider {
         this.analyzeAIResponse(generateResult);
         this.logGenerationComplete(generateResult);
         const responseTime = Date.now() - startTime;
-        await this.recordPerformanceMetrics(generateResult.usage, responseTime);
         const { toolsUsed, toolExecutions } = this.extractToolInformation(generateResult);
         // Prefer the per-call recorder's real records (params/result/timing per
         // execution); fall back to a conversion of the step-extraction entries
         // for tools the recorder could not wrap (provider-executed tools).
         const toolExecutionRecords = resolveToolExecutionRecords(options, toolExecutions);
         let enhancedResult = this.formatEnhancedResult(generateResult, tools, toolsUsed, toolExecutionRecords, options);
+        // Recorded AFTER formatEnhancedResult so telemetry sees the same usage
+        // the caller gets: the cross-step aggregate (totalUsage, not last-step
+        // usage) WITH the providerMetadata cache merge applied — otherwise
+        // providers whose cache data lives only in providerMetadata would have
+        // their cache tokens billed at the full input rate in OTEL metrics,
+        // diverging from analytics.cost.
+        await this.recordPerformanceMetrics(enhancedResult.usage, responseTime);
         enhancedResult = await this.synthesizeAIResponseIfNeeded(enhancedResult, options);
         const finalResult = await this.enhanceResult(enhancedResult, options, startTime);
         return finalResult;

package/dist/lib/core/modules/GenerationHandler.d.ts

@@ -68,6 +68,14 @@ export declare class GenerationHandler {
      * The AI SDK's LanguageModelUsage only has inputTokens/outputTokens.
      * Cache metrics are surfaced via providerMetadata by provider-specific SDK adapters.
      */
+    /**
+     * Set gen_ai usage attributes + cache-aware cost on the span from the
+     * CROSS-STEP aggregate (result.totalUsage). result.usage is the LAST step
+     * only — using it undercounted every multi-step tool loop, and pricing the
+     * raw cache-inclusive inputTokens without the cache fields billed cache
+     * reads at the full input rate.
+     */
+    private setUsageSpanAttributes;
     private extractCacheMetricsFromProviderMetadata;
     /**
      * Log generation completion information

package/dist/lib/core/modules/GenerationHandler.js

@@ -475,18 +475,7 @@ export class GenerationHandler {
                     });
                 }
                 // Set token usage and completion attributes on span
-                if (result.usage) {
-                    span.setAttribute("gen_ai.usage.input_tokens", result.usage.inputTokens || 0);
-                    span.setAttribute("gen_ai.usage.output_tokens", result.usage.outputTokens || 0);
-                    // Cost on span so users can query "what did this trace cost?"
-                    const cost = calculateCost(this.providerName, this.modelName, {
-                        input: result.usage.inputTokens || 0,
-                        output: result.usage.outputTokens || 0,
-                        total: (result.usage.inputTokens || 0) +
-                            (result.usage.outputTokens || 0),
-                    });
-                    span.setAttribute("neurolink.cost", cost ?? 0);
-                }
+                this.setUsageSpanAttributes(span, result);
                 if (result.finishReason) {
                     span.setAttribute("gen_ai.response.finish_reason", result.finishReason);
                 }
@@ -559,17 +548,7 @@ export class GenerationHandler {
                         toolCallsTotal: result.toolCalls?.length || 0,
                         responseChars: result.text?.length || 0,
                     });
-                    if (result.usage) {
-                        span.setAttribute("gen_ai.usage.input_tokens", result.usage.inputTokens || 0);
-                        span.setAttribute("gen_ai.usage.output_tokens", result.usage.outputTokens || 0);
-                        const fallbackCost = calculateCost(this.providerName, this.modelName, {
-                            input: result.usage.inputTokens || 0,
-                            output: result.usage.outputTokens || 0,
-                            total: (result.usage.inputTokens || 0) +
-                                (result.usage.outputTokens || 0),
-                        });
-                        span.setAttribute("neurolink.cost", fallbackCost ?? 0);
-                    }
+                    this.setUsageSpanAttributes(span, result);
                     if (result.finishReason) {
                         span.setAttribute("gen_ai.response.finish_reason", result.finishReason);
                     }
@@ -606,17 +585,7 @@ export class GenerationHandler {
                         "retry.strategy": "temperature_omitted",
                     });
                     span.setAttribute("retry.count", 1);
-                    if (result.usage) {
-                        span.setAttribute("gen_ai.usage.input_tokens", result.usage.inputTokens || 0);
-                        span.setAttribute("gen_ai.usage.output_tokens", result.usage.outputTokens || 0);
-                        const noTempCost = calculateCost(this.providerName, this.modelName, {
-                            input: result.usage.inputTokens || 0,
-                            output: result.usage.outputTokens || 0,
-                            total: (result.usage.inputTokens || 0) +
-                                (result.usage.outputTokens || 0),
-                        });
-                        span.setAttribute("neurolink.cost", noTempCost ?? 0);
-                    }
+                    this.setUsageSpanAttributes(span, result);
                     if (result.finishReason) {
                         span.setAttribute("gen_ai.response.finish_reason", result.finishReason);
                     }
@@ -640,6 +609,26 @@ export class GenerationHandler {
      * The AI SDK's LanguageModelUsage only has inputTokens/outputTokens.
      * Cache metrics are surfaced via providerMetadata by provider-specific SDK adapters.
      */
+    /**
+     * Set gen_ai usage attributes + cache-aware cost on the span from the
+     * CROSS-STEP aggregate (result.totalUsage). result.usage is the LAST step
+     * only — using it undercounted every multi-step tool loop, and pricing the
+     * raw cache-inclusive inputTokens without the cache fields billed cache
+     * reads at the full input rate.
+     */
+    setUsageSpanAttributes(span, result) {
+        const aggregate = result.totalUsage ?? result.usage;
+        if (!aggregate) {
+            return;
+        }
+        span.setAttribute("gen_ai.usage.input_tokens", aggregate.inputTokens || 0);
+        span.setAttribute("gen_ai.usage.output_tokens", aggregate.outputTokens || 0);
+        // Cost on span so users can query "what did this trace cost?" —
+        // extractTokenUsage rebases input onto the uncached remainder and
+        // surfaces the cache fields so calculateCost prices each tier.
+        const cost = calculateCost(this.providerName, this.modelName, extractTokenUsage(aggregate));
+        span.setAttribute("neurolink.cost", cost ?? 0);
+    }
     extractCacheMetricsFromProviderMetadata(generateResult) {
         const providerMeta = generateResult.providerMetadata;
         if (!providerMeta) {
@@ -841,10 +830,11 @@ export class GenerationHandler {
             jsonTruncated = true;
             logger.warn("[GenerationHandler] Structured output truncated by token cap (finishReason=length); increase maxTokens", { provider: this.providerName, model: this.modelName });
         }
-        // Extract usage with support for different formats and reasoning tokens
-        // Note: The AI SDK bundles thinking tokens into promptTokens for Google models.
-        // Separate reasoningTokens tracking will work when/if the AI SDK adds support.
-        const usage = extractTokenUsage(generateResult.usage);
+        // Extract usage with support for different formats and reasoning tokens.
+        // totalUsage is the CROSS-STEP aggregate; generateResult.usage is the
+        // LAST step only, which silently dropped every prior step of a
+        // multi-step tool loop.
+        const usage = extractTokenUsage(generateResult.totalUsage ?? generateResult.usage);
         // Merge cache metrics from providerMetadata if not already present in usage
         // The AI SDK's LanguageModelUsage doesn't include cache tokens; they come from
         // provider-specific metadata (e.g. Anthropic's providerMetadata.anthropic)

package/dist/lib/core/modules/TelemetryHandler.d.ts

@@ -14,7 +14,7 @@
  * @module core/modules/TelemetryHandler
  */
 import type { NeuroLink } from "../../neurolink.js";
-import type { StreamOptions, AIProviderName, AnalyticsData, EnhancedGenerateResult, EvaluationData, TextGenerationOptions } from "../../types/index.js";
+import type { StreamOptions, AIProviderName, AnalyticsData, EnhancedGenerateResult, EvaluationData, RawUsageObject, TextGenerationOptions, TokenUsage } from "../../types/index.js";
 /**
  * TelemetryHandler class - Handles analytics and telemetry for AI providers
  */
@@ -34,10 +34,7 @@ export declare class TelemetryHandler {
     /**
      * Record performance metrics for a generation
      */
-    recordPerformanceMetrics(usage: {
-        inputTokens: number | undefined;
-        outputTokens: number | undefined;
-    } | undefined, responseTime: number): Promise<void>;
+    recordPerformanceMetrics(usage: RawUsageObject | undefined, responseTime: number): Promise<void>;
     /**
      * Calculate actual cost based on token usage and provider configuration.
      *
@@ -50,10 +47,7 @@ export declare class TelemetryHandler {
      * causing a ~1,780x under-estimate when the actual model was Claude Sonnet
      * on Vertex AI ($0.000060 vs $0.106895 for the same request).
      */
-    calculateActualCost(usage: {
-        inputTokens?: number | undefined;
-        outputTokens?: number | undefined;
-    }): Promise<number>;
+    calculateActualCost(usage: TokenUsage): Promise<number>;
     /**
      * Create telemetry configuration for Vercel AI SDK experimental_telemetry
      * This enables automatic OpenTelemetry tracing when telemetry is enabled

package/dist/lib/core/modules/TelemetryHandler.js

@@ -14,6 +14,7 @@
  * @module core/modules/TelemetryHandler
  */
 import { nanoid } from "nanoid";
+import { extractTokenUsage } from "../../utils/tokenUtils.js";
 import { logger } from "../../utils/logger.js";
 import { recordProviderPerformanceFromMetrics } from "../evaluationProviders.js";
 import { modelConfig } from "../modelConfiguration.js";
@@ -75,8 +76,12 @@ export class TelemetryHandler {
      */
     async recordPerformanceMetrics(usage, responseTime) {
         try {
-            const totalTokens = (usage?.inputTokens || 0) + (usage?.outputTokens || 0);
-            const actualCost = await this.calculateActualCost(usage || { inputTokens: 0, outputTokens: 0 });
+            // Normalize first: rebases cache-inclusive ai@6 input onto the uncached
+            // remainder and surfaces cache tiers so the cost is priced per tier
+            // instead of billing cache reads at the full input rate.
+            const tokenUsage = extractTokenUsage(usage);
+            const totalTokens = tokenUsage.total;
+            const actualCost = await this.calculateActualCost(tokenUsage);
             recordProviderPerformanceFromMetrics(this.providerName, {
                 responseTime,
                 tokensGenerated: totalTokens,
@@ -109,25 +114,25 @@ export class TelemetryHandler {
      */
     async calculateActualCost(usage) {
         try {
-            const promptTokens = usage?.inputTokens || 0;
-            const completionTokens = usage?.outputTokens || 0;
             // Try the per-model pricing table first (includes correct rates for
             // Claude on Vertex, cache token rates, etc.)
             if (hasPricing(this.providerName, this.modelName)) {
-                return calculateCost(this.providerName, this.modelName, {
-                    input: promptTokens,
-                    output: completionTokens,
-                    total: promptTokens + completionTokens,
-                });
+                return calculateCost(this.providerName, this.modelName, usage);
             }
             // Fall back to provider-level default cost from configuration system
             const costInfo = modelConfig.getCostInfo(this.providerName, this.modelName);
             if (!costInfo) {
                 return 0; // No cost info available
             }
-            // Calculate cost per 1K tokens
-            const inputCost = (promptTokens / 1000) * costInfo.input;
-            const outputCost = (completionTokens / 1000) * costInfo.output;
+            // Calculate cost per 1K tokens. costInfo has no cache tiers, so cache
+            // tokens are billed at the input rate — the same total a provider
+            // without cache-aware splitting would have reported, never $0.
+            const inputCost = ((usage.input +
+                (usage.cacheReadTokens ?? 0) +
+                (usage.cacheCreationTokens ?? 0)) /
+                1000) *
+                costInfo.input;
+            const outputCost = (usage.output / 1000) * costInfo.output;
             return inputCost + outputCost;
         }
         catch (error) {

package/dist/lib/neurolink.js

@@ -2746,7 +2746,10 @@ Current user's request: ${currentInput}`;
         if (!usage || model === "unknown") {
             return undefined;
         }
-        const totalCost = calculateAdvancedCost(model, usage.input || 0, usage.output || 0, provider === "unknown" ? undefined : provider);
+        const totalCost = calculateAdvancedCost(model, usage.input || 0, usage.output || 0, provider === "unknown" ? undefined : provider, {
+            cacheReadTokens: usage.cacheReadTokens,
+            cacheCreationTokens: usage.cacheCreationTokens,
+        });
         return totalCost > 0 ? totalCost : undefined;
     }
     /**
@@ -4134,6 +4137,20 @@ Current user's request: ${currentInput}`;
                     input: textResult.usage.input || 0,
                     output: textResult.usage.output || 0,
                     total: textResult.usage.total || 0,
+                    // Optional tiers must be forwarded or they silently vanish at
+                    // this DTO boundary (cache-aware cost + savings need them).
+                    ...(textResult.usage.cacheReadTokens !== undefined && {
+                        cacheReadTokens: textResult.usage.cacheReadTokens,
+                    }),
+                    ...(textResult.usage.cacheCreationTokens !== undefined && {
+                        cacheCreationTokens: textResult.usage.cacheCreationTokens,
+                    }),
+                    ...(textResult.usage.reasoning !== undefined && {
+                        reasoning: textResult.usage.reasoning,
+                    }),
+                    ...(textResult.usage.cacheSavingsPercent !== undefined && {
+                        cacheSavingsPercent: textResult.usage.cacheSavingsPercent,
+                    }),
                 }
                 : undefined,
             responseTime: textResult.responseTime,