diff --git a/README.md b/README.md
index 883a5f1..ec3bf15 100644
--- a/README.md
+++ b/README.md
@@ -123,7 +123,7 @@ score is the weighted mean of the scorer scores.
 | `embedding-similarity` | cosine similarity >= threshold | `expected`, `threshold` |
 | `llm-judge` | a judge model scores >= threshold | `criteria`, `expected`, `threshold`, `model` |
 | `latency` | call latency within budget | `budgetMs` |
-| `cost` | estimated call cost within budget | `budgetUsd` |
+| `cost` | estimated call cost within budget | `budgetUsd` (provider must set `costPer1kTokens` so `costUsd` is reported) |
 | `rubric` | weighted criteria score >= threshold | `criteria[]`, `threshold` |
 
 Two scorers are pluggable and ship with **deterministic offline fallbacks** so tests and
diff --git a/src/scorers/cost.ts b/src/scorers/cost.ts
index 5458d86..5cdd977 100644
--- a/src/scorers/cost.ts
+++ b/src/scorers/cost.ts
@@ -6,6 +6,9 @@ import { result } from "./util.js";
  *
  * Options:
  *  - `budgetUsd`: the maximum acceptable cost in USD (required)
+ *
+ * Requires the provider to report `costUsd` (configure `costPer1kTokens` on OpenAI-compatible
+ * and Anthropic providers).
  */
 export const costScorer: Scorer = {
   type: "cost",
@@ -14,7 +17,14 @@ export const costScorer: Scorer = {
     if (!Number.isFinite(budgetUsd) || budgetUsd <= 0) {
       return result(spec, { score: 0, passed: false, reason: "invalid or missing budgetUsd" });
     }
-    const actual = ctx.response.costUsd ?? 0;
+    if (ctx.response.costUsd === undefined) {
+      return result(spec, {
+        score: 0,
+        passed: false,
+        reason: "provider reported no cost (set costPer1kTokens on the provider)",
+      });
+    }
+    const actual = ctx.response.costUsd;
     const passed = actual <= budgetUsd;
     const score = actual <= budgetUsd ? 1 : Math.max(0, 1 - (actual - budgetUsd) / budgetUsd);
     return result(spec, {
diff --git a/tests/scorers.test.ts b/tests/scorers.test.ts
index 78bf373..e7f15e5 100644
--- a/tests/scorers.test.ts
+++ b/tests/scorers.test.ts
@@ -176,6 +176,12 @@ describe("latency and cost budgets", () => {
     const r = await run(costScorer, { type: "cost", budgetUsd: 0.00001 }, ctx("x", { costUsd: 0.001 }));
     expect(r.passed).toBe(false);
   });
+  it("cost fails when provider reported no cost", async () => {
+    const r = await run(costScorer, { type: "cost", budgetUsd: 1 }, ctx("x", { costUsd: undefined }));
+    expect(r.passed).toBe(false);
+    expect(r.score).toBe(0);
+    expect(r.reason).toMatch(/no cost/i);
+  });
 });
 
 describe("rubric", () => {