jshah-do · Oct 1, 2026 · Sep 30, 2026 · Sep 30, 2026 · Sep 30, 2026
diff --git a/src/benchmarks/agent-cli/harness.ts b/src/benchmarks/agent-cli/harness.ts
@@ -227,10 +227,11 @@ function reasoningFromContent(content: unknown): string | undefined {
 
 function usageFromResult(result: Record<string, unknown>): ModelUsage {
   const { usage } = result;
+  const cacheReadTokens = numberField(usage, "cache_read_input_tokens");
   const inputTokens =
     numberField(usage, "input_tokens") +
     numberField(usage, "cache_creation_input_tokens") +
-    numberField(usage, "cache_read_input_tokens");
+    cacheReadTokens;
   const outputTokens = numberField(usage, "output_tokens");
   const details = isRecord(usage) ? usage["output_tokens_details"] : undefined;
   const reasoningTokens = numberField(details, "thinking_tokens");
@@ -243,6 +244,7 @@ function usageFromResult(result: Record<string, unknown>): ModelUsage {
     inputTokens,
     outputTokens,
     totalTokens: inputTokens + outputTokens,
+    cacheReadTokens,
     reasoningTokens,
     totalCost: numberField(result, "total_cost_usd"),
     ...definedValues({
@@ -704,6 +706,7 @@ function parseJsonAgentStream(stdout: string): OriAgentRun {
           inputTokens: inputTokens + cacheRead + cacheWrite,
           outputTokens,
           totalTokens,
+          cacheReadTokens: cacheRead,
           reasoningTokens,
           totalCost,
         }

diff --git a/src/benchmarks/deep-swe/solver.ts b/src/benchmarks/deep-swe/solver.ts
@@ -472,6 +472,7 @@ const ZERO_AGENT_USAGE: ModelUsage = {
   inputTokens: 0,
   outputTokens: 0,
   totalTokens: 0,
+  cacheReadTokens: 0,
   reasoningTokens: 0,
   totalCost: 0,
 };
diff --git a/src/benchmarks/draco/benchmark.test.ts b/src/benchmarks/draco/benchmark.test.ts
@@ -35,6 +35,7 @@ function runResultWith(
       inputTokens: 0,
       outputTokens: 0,
       totalTokens: 0,
+      cacheReadTokens: 0,
       reasoningTokens: 0,
       totalCost: 0,
       generationTimeMs: 0,

diff --git a/src/benchmarks/harbor/agent-loop.ts b/src/benchmarks/harbor/agent-loop.ts
@@ -438,6 +438,7 @@ export interface UsageAccumulator {
   inputTokens: number;
   outputTokens: number;
   totalTokens: number;
+  cacheReadTokens: number;
   reasoningTokens: number;
   totalCost: number;
   webSearchRequests: number;
@@ -451,6 +452,7 @@ function newUsageAccumulator(): UsageAccumulator {
     inputTokens: 0,
     outputTokens: 0,
     totalTokens: 0,
+    cacheReadTokens: 0,
     reasoningTokens: 0,
     totalCost: 0,
     webSearchRequests: 0,
@@ -467,6 +469,7 @@ function addUsage(acc: UsageAccumulator, usage: ModelUsage | undefined): void {
   acc.inputTokens += usage.inputTokens ?? 0;
   acc.outputTokens += usage.outputTokens ?? 0;
   acc.totalTokens += usage.totalTokens ?? 0;
+  acc.cacheReadTokens += usage.cacheReadTokens ?? 0;
   acc.reasoningTokens += usage.reasoningTokens ?? 0;
   acc.totalCost += usage.totalCost ?? 0;
   const serverToolUse = usage.serverToolUse;
@@ -492,6 +495,7 @@ function toModelUsage(acc: UsageAccumulator): ModelUsage {
     inputTokens: acc.inputTokens,
     outputTokens: acc.outputTokens,
     totalTokens: acc.totalTokens,
+    cacheReadTokens: acc.cacheReadTokens,
     reasoningTokens: acc.reasoningTokens,
     totalCost: acc.totalCost,
     serverToolUse: acc.seenServerToolUse

diff --git a/src/benchmarks/mmlu-pro.test.ts b/src/benchmarks/mmlu-pro.test.ts
@@ -47,6 +47,7 @@ function runResultWith(
       inputTokens: 0,
       outputTokens: 0,
       totalTokens: 0,
+      cacheReadTokens: 0,
       reasoningTokens: 0,
       totalCost: 0,
       generationTimeMs: 0,
@@ -144,6 +145,7 @@ describe("MMLU-Pro runLevelScores", () => {
         inputTokens: 0,
         outputTokens: 0,
         totalTokens: 0,
+        cacheReadTokens: 0,
         reasoningTokens: 0,
         totalCost: 0,
         generationTimeMs: 0,

diff --git a/src/benchmarks/search/browsecomp/benchmark.test.ts b/src/benchmarks/search/browsecomp/benchmark.test.ts
@@ -308,6 +308,7 @@ describe("browseCompRunLevelScores", () => {
         inputTokens: 0,
         outputTokens: 0,
         totalTokens: 0,
+        cacheReadTokens: 0,
         reasoningTokens: 0,
         totalCost: 0,
         generationTimeMs: 0,

diff --git a/src/benchmarks/search/core/usage.test.ts b/src/benchmarks/search/core/usage.test.ts
@@ -9,6 +9,7 @@ describe("mergeModelUsages", () => {
           inputTokens: 10,
           outputTokens: 5,
           totalTokens: 15,
+          cacheReadTokens: 0,
           reasoningTokens: 2,
           totalCost: 0.02,
           serverToolUse: {
@@ -22,6 +23,7 @@ describe("mergeModelUsages", () => {
           inputTokens: 4,
           outputTokens: 2,
           totalTokens: 6,
+          cacheReadTokens: 0,
           reasoningTokens: 1,
           totalCost: 0.001,
           serverToolUse: {
@@ -36,6 +38,7 @@ describe("mergeModelUsages", () => {
       inputTokens: 14,
       outputTokens: 7,
       totalTokens: 24,
+      cacheReadTokens: 0,
       reasoningTokens: 3,
       totalCost: 0.021,
       serverToolUse: {
@@ -45,6 +48,17 @@ describe("mergeModelUsages", () => {
       },
     });
   });
+  it("sums cacheReadTokens when present", () => {
+    expect(
+      mergeModelUsages([
+        { inputTokens: 10, cacheReadTokens: 4 },
+        { inputTokens: 5, cacheReadTokens: 2 },
+      ])
+    ).toEqual({
+      inputTokens: 15,
+      cacheReadTokens: 6,
+    });
+  });
   it("returns undefined when no call reports usage", () => {
     expect(mergeModelUsages([undefined, undefined])).toBeUndefined();
   });

diff --git a/src/benchmarks/search/core/usage.ts b/src/benchmarks/search/core/usage.ts
@@ -43,6 +43,7 @@ export function mergeModelUsages(
   const inputTokens = sum(defined.map((usage) => usage.inputTokens));
   const outputTokens = sum(defined.map((usage) => usage.outputTokens));
   const totalTokens = sum(defined.map((usage) => usage.totalTokens));
+  const cacheReadTokens = sum(defined.map((usage) => usage.cacheReadTokens));
   const reasoningTokens = sum(defined.map((usage) => usage.reasoningTokens));
   const totalCost = sum(defined.map((usage) => usage.totalCost));
   const serverToolUse = mergeServerToolUse(
@@ -54,6 +55,7 @@ export function mergeModelUsages(
     inputTokens,
     outputTokens,
     totalTokens,
+    cacheReadTokens,
     reasoningTokens,
     totalCost,
     serverToolUse,

diff --git a/src/benchmarks/search/dsqa/benchmark.test.ts b/src/benchmarks/search/dsqa/benchmark.test.ts
@@ -232,6 +232,7 @@ describe("DSQA benchmark", () => {
         inputTokens: 0,
         outputTokens: 0,
         totalTokens: 0,
+        cacheReadTokens: 0,
         reasoningTokens: 0,
         totalCost: 0,
         generationTimeMs: 0,
@@ -334,6 +335,7 @@ describe("DSQA metrics", () => {
         inputTokens: 0,
         outputTokens: 0,
         totalTokens: 0,
+        cacheReadTokens: 0,
         reasoningTokens: 0,
         totalCost: 0,
         generationTimeMs: 0,

diff --git a/src/benchmarks/search/hle/benchmark.test.ts b/src/benchmarks/search/hle/benchmark.test.ts
@@ -293,6 +293,7 @@ function runResult(sampleScores: SampleScore[]): RunResult {
       inputTokens: 0,
       outputTokens: 0,
       totalTokens: 0,
+      cacheReadTokens: 0,
       reasoningTokens: 0,
       totalCost: 0,
       generationTimeMs: 0,

diff --git a/src/benchmarks/search/widesearch/benchmark.test.ts b/src/benchmarks/search/widesearch/benchmark.test.ts
@@ -242,6 +242,7 @@ describe("wideSearchScorer and run-level scores", () => {
         inputTokens: 0,
         outputTokens: 0,
         totalTokens: 0,
+        cacheReadTokens: 0,
         reasoningTokens: 0,
         totalCost: 0,
         generationTimeMs: 0,
@@ -268,6 +269,7 @@ describe("wideSearchScorer and run-level scores", () => {
         inputTokens: 0,
         outputTokens: 0,
         totalTokens: 0,
+        cacheReadTokens: 0,
         reasoningTokens: 0,
         totalCost: 0,
         generationTimeMs: 0,
@@ -338,6 +340,7 @@ describe("wideSearchScorer and run-level scores", () => {
         inputTokens: 0,
         outputTokens: 0,
         totalTokens: 0,
+        cacheReadTokens: 0,
         reasoningTokens: 0,
         totalCost: 0,
         generationTimeMs: 0,

diff --git a/src/benchmarks/swe-atlas/solver.ts b/src/benchmarks/swe-atlas/solver.ts
@@ -54,6 +54,7 @@ const ZERO_AGENT_USAGE: ModelUsage = {
   inputTokens: 0,
   outputTokens: 0,
   totalTokens: 0,
+  cacheReadTokens: 0,
   reasoningTokens: 0,
   totalCost: 0,
 };

diff --git a/src/benchmarks/swe-bench/solver.ts b/src/benchmarks/swe-bench/solver.ts
@@ -400,6 +400,7 @@ const ZERO_AGENT_USAGE: ModelUsage = {
   inputTokens: 0,
   outputTokens: 0,
   totalTokens: 0,
+  cacheReadTokens: 0,
   reasoningTokens: 0,
   totalCost: 0,
 };
diff --git a/src/benchmarks/tau-bench-airline/solver.ts b/src/benchmarks/tau-bench-airline/solver.ts
@@ -106,6 +106,7 @@ export function airlineSolver({
         inputTokens: 0,
         outputTokens: 0,
         totalTokens: 0,
+        cacheReadTokens: 0,
         reasoningTokens: 0,
         totalCost: 0,
       };
@@ -157,6 +158,7 @@ export function airlineSolver({
                 accUsage.inputTokens += output.usage.inputTokens ?? 0;
                 accUsage.outputTokens += output.usage.outputTokens ?? 0;
                 accUsage.totalTokens += output.usage.totalTokens ?? 0;
+                accUsage.cacheReadTokens += output.usage.cacheReadTokens ?? 0;
                 accUsage.reasoningTokens += output.usage.reasoningTokens ?? 0;
                 accUsage.totalCost += output.usage.totalCost ?? 0;
               }
@@ -205,6 +207,7 @@ export function airlineSolver({
         inputTokens: accUsage.inputTokens,
         outputTokens: accUsage.outputTokens,
         totalTokens: accUsage.totalTokens,
+        cacheReadTokens: accUsage.cacheReadTokens,
         reasoningTokens: accUsage.reasoningTokens,
         totalCost: accUsage.totalCost,
       };

diff --git a/src/benchmarks/tau3-bench-banking/solver.ts b/src/benchmarks/tau3-bench-banking/solver.ts
@@ -226,6 +226,7 @@ export function bankingSolver({
         inputTokens: 0,
         outputTokens: 0,
         totalTokens: 0,
+        cacheReadTokens: 0,
         reasoningTokens: 0,
         totalCost: 0,
       };
@@ -299,6 +300,7 @@ export function bankingSolver({
               accUsage.inputTokens += output.usage.inputTokens ?? 0;
               accUsage.outputTokens += output.usage.outputTokens ?? 0;
               accUsage.totalTokens += output.usage.totalTokens ?? 0;
+              accUsage.cacheReadTokens += output.usage.cacheReadTokens ?? 0;
               accUsage.reasoningTokens += output.usage.reasoningTokens ?? 0;
               accUsage.totalCost += output.usage.totalCost ?? 0;
             }
@@ -341,6 +343,7 @@ export function bankingSolver({
         inputTokens: accUsage.inputTokens,
         outputTokens: accUsage.outputTokens,
         totalTokens: accUsage.totalTokens,
+        cacheReadTokens: accUsage.cacheReadTokens,
         reasoningTokens: accUsage.reasoningTokens,
         totalCost: accUsage.totalCost,
       };

diff --git a/src/benchmarks/terminal-bench/ori-parquet.test.ts b/src/benchmarks/terminal-bench/ori-parquet.test.ts
@@ -134,6 +134,7 @@ async function runOriSampleToParquetRow() {
         inputTokens: usage?.inputTokens ?? 0,
         outputTokens: usage?.outputTokens ?? 0,
         totalTokens: usage?.totalTokens ?? 0,
+        cacheReadTokens: usage?.cacheReadTokens ?? 0,
         reasoningTokens: usage?.reasoningTokens ?? 0,
         totalCost: usage?.totalCost ?? 0,
         generationTimeMs: state.output?.generationTimeMs ?? 0,
@@ -144,6 +145,7 @@ async function runOriSampleToParquetRow() {
           epoch: 0,
           score,
           messages: state.messages,
+          ...(usage !== undefined && { usage }),
           ...(state.responseItems !== undefined && {
             responseItems: state.responseItems,
           }),

diff --git a/src/benchmarks/terminal-bench/ori-solver.test.ts b/src/benchmarks/terminal-bench/ori-solver.test.ts
@@ -1274,6 +1274,7 @@ describe("terminal-bench Prime Agent via ori", () => {
       inputTokens: 4494,
       outputTokens: 3,
       totalTokens: 4497,
+      cacheReadTokens: 0,
       reasoningTokens: 0,
       totalCost: 0.0018023999999999998,
     });
@@ -1540,6 +1541,7 @@ describe("terminal-bench Prime Agent via ori", () => {
       inputTokens: 37,
       outputTokens: 8,
       totalTokens: 126,
+      cacheReadTokens: 10,
       reasoningTokens: 0,
       totalCost: 0,
     });
@@ -1619,6 +1621,7 @@ describe("terminal-bench omp via ori", () => {
           cacheRead: 0,
           cacheWrite: 0,
           totalTokens: 127,
+          cacheReadTokens: 0,
           reasoningTokens: 4,
           cost: {
             input: 0.000048,
@@ -1696,6 +1699,7 @@ describe("terminal-bench omp via ori", () => {
       inputTokens: 120,
       outputTokens: 7,
       totalTokens: 127,
+      cacheReadTokens: 0,
       reasoningTokens: 4,
       totalCost: 0.0000536,
     });

diff --git a/src/benchmarks/terminal-bench/ori-solver.ts b/src/benchmarks/terminal-bench/ori-solver.ts
@@ -26,6 +26,7 @@ const ZERO_USAGE: ModelUsage = {
   inputTokens: 0,
   outputTokens: 0,
   totalTokens: 0,
+  cacheReadTokens: 0,
   reasoningTokens: 0,
   totalCost: 0,
 };

diff --git a/src/benchmarks/vgi-bench/benchmark.test.ts b/src/benchmarks/vgi-bench/benchmark.test.ts
@@ -375,6 +375,7 @@ describe("VGI-Bench registry", () => {
         inputTokens: 0,
         outputTokens: 0,
         totalTokens: 0,
+        cacheReadTokens: 0,
         reasoningTokens: 0,
         totalCost: 0,
         generationTimeMs: 0,

diff --git a/src/benchmarks/wandr/scorer.test.ts b/src/benchmarks/wandr/scorer.test.ts
@@ -29,6 +29,7 @@ function resultWith(values: readonly number[]): RunResult {
       inputTokens: 0,
       outputTokens: 0,
       totalTokens: 0,
+      cacheReadTokens: 0,
       reasoningTokens: 0,
       totalCost: 0,
       generationTimeMs: 0,

diff --git a/src/harness/core.ts b/src/harness/core.ts
@@ -136,6 +136,7 @@ export interface UsageTotals {
   readonly inputTokens: number;
   readonly outputTokens: number;
   readonly totalTokens: number;
+  readonly cacheReadTokens: number;
   readonly reasoningTokens: number;
   readonly totalCost: number;
   readonly generationTimeMs: number;

diff --git a/src/harness/metric.ts b/src/harness/metric.ts
@@ -1,4 +1,10 @@
-import type { ModelMessage, ResponseItem, Score, ScoreValue } from "./core";
+import type {
+  ModelMessage,
+  ModelUsage,
+  ResponseItem,
+  Score,
+  ScoreValue,
+} from "./core";
 import { ScoreValue as SV, scoreToNumber } from "./core";
 
 export interface SampleScore {
@@ -10,6 +16,7 @@ export interface SampleScore {
   readonly requestBody?: Readonly<Record<string, unknown>>;
   readonly generationIds?: readonly string[];
   readonly generationTimeMs?: number;
+  readonly usage?: ModelUsage;
   readonly metadata?: Readonly<Record<string, unknown>>;
   readonly input?: string;
   readonly target?: string;

diff --git a/src/harness/progress.ts b/src/harness/progress.ts
@@ -129,6 +129,7 @@ export const CheckpointDataSchema = z.object({
       inputTokens: z.number(),
       outputTokens: z.number(),
       totalTokens: z.number(),
+      cacheReadTokens: z.number().optional(),
       reasoningTokens: z.number(),
       totalCost: z.number(),
       webSearchRequests: z.number().optional(),