diff --git a/.changeset/rich-token-usage.md b/.changeset/rich-token-usage.md
new file mode 100644
index 0000000..20d33be
--- /dev/null
+++ b/.changeset/rich-token-usage.md
@@ -0,0 +1,5 @@
+---
+"@loveholidays/eval-kit": minor
+---
+
+Expose detailed AI SDK token usage on evaluator results.
diff --git a/src/evaluators/evaluator.spec.ts b/src/evaluators/evaluator.spec.ts
index 0f5b562..786689b 100644
--- a/src/evaluators/evaluator.spec.ts
+++ b/src/evaluators/evaluator.spec.ts
@@ -85,6 +85,57 @@ describe("Evaluator", () => {
 			});
 		});
 
+		it("should preserve detailed token usage", async () => {
+			const model = createMockModel();
+
+			mockGenerateText.mockResolvedValue({
+				output: { score: 90, feedback: "Excellent fluency" },
+				usage: {
+					inputTokens: 150,
+					inputTokenDetails: {
+						noCacheTokens: 100,
+						cacheReadTokens: 40,
+						cacheWriteTokens: 10,
+					},
+					outputTokens: 25,
+					outputTokenDetails: {
+						textTokens: 15,
+						reasoningTokens: 10,
+					},
+					totalTokens: 175,
+					reasoningTokens: 10,
+					cachedInputTokens: 40,
+				},
+			});
+
+			const evaluator = new Evaluator({
+				name: "fluency",
+				model,
+				evaluationPrompt: "Rate the fluency of: {{candidateText}}",
+			});
+
+			const result = await evaluator.evaluate({
+				candidateText: "The quick brown fox jumps over the lazy dog.",
+			});
+
+			expect(result.processingStats.tokenUsage).toEqual({
+				inputTokens: 150,
+				inputTokenDetails: {
+					noCacheTokens: 100,
+					cacheReadTokens: 40,
+					cacheWriteTokens: 10,
+				},
+				outputTokens: 25,
+				outputTokenDetails: {
+					textTokens: 15,
+					reasoningTokens: 10,
+				},
+				totalTokens: 175,
+				reasoningTokens: 10,
+				cachedInputTokens: 40,
+			});
+		});
+
 		it("should evaluate with reference text", async () => {
 			const model = createMockModel();
 
diff --git a/src/evaluators/evaluator.ts b/src/evaluators/evaluator.ts
index 09f0771..9cd81e7 100644
--- a/src/evaluators/evaluator.ts
+++ b/src/evaluators/evaluator.ts
@@ -10,9 +10,24 @@ import type {
 	EvaluationInput,
 	EvaluatorConfig,
 	EvaluatorResult,
+	TokenUsage,
 } from "../types/evaluator.js";
 import { TemplateRenderer } from "../utils/template-engine.js";
 
+function getNumber(value: unknown): number | undefined {
+	return typeof value === "number" ? value : undefined;
+}
+
+function getRecord(value: unknown): Record<string, unknown> | undefined {
+	return value && typeof value === "object"
+		? (value as Record<string, unknown>)
+		: undefined;
+}
+
+function hasNumberValue(value: Record<string, unknown>): boolean {
+	return Object.values(value).some((entry) => typeof entry === "number");
+}
+
 export class Evaluator {
 	readonly name: string;
 	readonly timeout?: number;
@@ -171,9 +186,7 @@ export class Evaluator {
 
 	private setTokenAttributes(
 		span: EvalKitSpan,
-		tokenUsage:
-			| { inputTokens?: number; outputTokens?: number; totalTokens?: number }
-			| undefined,
+		tokenUsage: TokenUsage | undefined,
 	): void {
 		if (tokenUsage?.inputTokens !== undefined) {
 			span.setAttribute(
@@ -207,22 +220,51 @@ export class Evaluator {
 		};
 	}
 
-	private extractTokenUsage(
-		usage: unknown,
-	):
-		| { inputTokens?: number; outputTokens?: number; totalTokens?: number }
-		| undefined {
+	private extractTokenUsage(usage: unknown): TokenUsage | undefined {
 		if (!usage) return undefined;
+		const usageRecord = getRecord(usage);
+		if (!usageRecord) return undefined;
+
+		const inputTokenDetails = this.extractInputTokenDetails(
+			usageRecord.inputTokenDetails,
+		);
+		const outputTokenDetails = this.extractOutputTokenDetails(
+			usageRecord.outputTokenDetails,
+		);
+
+		return {
+			inputTokens: getNumber(usageRecord.inputTokens),
+			...(inputTokenDetails ? { inputTokenDetails } : {}),
+			outputTokens: getNumber(usageRecord.outputTokens),
+			...(outputTokenDetails ? { outputTokenDetails } : {}),
+			totalTokens: getNumber(usageRecord.totalTokens),
+			reasoningTokens: getNumber(usageRecord.reasoningTokens),
+			cachedInputTokens: getNumber(usageRecord.cachedInputTokens),
+		};
+	}
+
+	private extractInputTokenDetails(
+		details: unknown,
+	): TokenUsage["inputTokenDetails"] | undefined {
+		const detailRecord = getRecord(details);
+		if (!detailRecord || !hasNumberValue(detailRecord)) return undefined;
+
+		return {
+			noCacheTokens: getNumber(detailRecord.noCacheTokens),
+			cacheReadTokens: getNumber(detailRecord.cacheReadTokens),
+			cacheWriteTokens: getNumber(detailRecord.cacheWriteTokens),
+		};
+	}
+
+	private extractOutputTokenDetails(
+		details: unknown,
+	): TokenUsage["outputTokenDetails"] | undefined {
+		const detailRecord = getRecord(details);
+		if (!detailRecord || !hasNumberValue(detailRecord)) return undefined;
+
 		return {
-			inputTokens: (usage as Record<string, unknown>).inputTokens as
-				| number
-				| undefined,
-			outputTokens: (usage as Record<string, unknown>).outputTokens as
-				| number
-				| undefined,
-			totalTokens: (usage as Record<string, unknown>).totalTokens as
-				| number
-				| undefined,
+			textTokens: getNumber(detailRecord.textTokens),
+			reasoningTokens: getNumber(detailRecord.reasoningTokens),
 		};
 	}
 
diff --git a/src/types/evaluator.ts b/src/types/evaluator.ts
index f0d71d9..34370a9 100644
--- a/src/types/evaluator.ts
+++ b/src/types/evaluator.ts
@@ -2,8 +2,19 @@ import type { LanguageModel } from "ai";
 
 export interface TokenUsage {
 	readonly inputTokens?: number;
+	readonly inputTokenDetails?: {
+		readonly noCacheTokens?: number;
+		readonly cacheReadTokens?: number;
+		readonly cacheWriteTokens?: number;
+	};
 	readonly outputTokens?: number;
+	readonly outputTokenDetails?: {
+		readonly textTokens?: number;
+		readonly reasoningTokens?: number;
+	};
 	readonly totalTokens?: number;
+	readonly reasoningTokens?: number;
+	readonly cachedInputTokens?: number;
 }
 
 export interface ProcessingStats {