diff --git a/services/gateway/src/usage.ts b/services/gateway/src/usage.ts
index 241a54c9..eb777579 100644
--- a/services/gateway/src/usage.ts
+++ b/services/gateway/src/usage.ts
@@ -112,12 +112,15 @@ export class UsageTee extends Transform {
       const parsed = JSON.parse(text);
       if (this.provider === "anthropic") {
         // Anthropic splits streamed usage across frames: `message_start` carries
-        // input and cache tokens under `message.usage`, `message_delta` carries
-        // output tokens under `delta.usage`, and a non-streamed body reports
-        // everything under a top-level `usage`. Merge each source separately
-        // rather than spreading them into one object: a spread copies absent
-        // keys as `undefined` and would drop an earlier frame's input tokens
-        // when the later frame reports only output.
+        // input and cache tokens under `message.usage`, and `message_delta`
+        // reports output tokens in a top-level `usage` — a sibling of `delta`,
+        // which itself carries only the stop fields. A non-streamed body also
+        // reports everything under a top-level `usage`. Merge each source
+        // separately rather than spreading them into one object: a spread
+        // copies absent keys as `undefined` and would drop an earlier frame's
+        // input tokens when the later frame reports only output. The
+        // `parsed.delta` source stays as tolerance for relays that nest usage
+        // inside the delta.
         for (const source of [parsed, parsed?.message, parsed?.delta]) {
           mergeUsage(this.usage, usageFromJson("anthropic", source));
         }
diff --git a/services/gateway/test/gateway.test.ts b/services/gateway/test/gateway.test.ts
index 6a8574a1..cda30409 100644
--- a/services/gateway/test/gateway.test.ts
+++ b/services/gateway/test/gateway.test.ts
@@ -1310,7 +1310,7 @@ async function buildStub(state: StubState) {
       });
       reply.raw.writeHead(200, { "content-type": "text/event-stream" });
       reply.raw.write(
-        'event: message_delta\ndata: {"type":"message_delta","delta":{"usage":{"output_tokens":333}}}\n\n',
+        'event: message_delta\ndata: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"output_tokens":333}}\n\n',
       );
       await new Promise((resolve) => setTimeout(resolve, 10_000));
       return reply;
@@ -1375,8 +1375,11 @@ async function buildStub(state: StubState) {
 
 function anthropicSseBody() {
   return [
-    'event: message_start\ndata: {"type":"message_start","message":{"usage":{"input_tokens":1000000}}}\n\n',
+    'event: message_start\ndata: {"type":"message_start","message":{"id":"msg_stub","type":"message","role":"assistant","model":"claude-sonnet-5","content":[],"stop_reason":null,"usage":{"input_tokens":1000000,"output_tokens":1}}}\n\n',
     'event: content_block_delta\ndata: {"type":"content_block_delta","delta":{"text":"ok"}}\n\n',
-    'event: message_delta\ndata: {"type":"message_delta","delta":{"usage":{"output_tokens":1000000}}}\n\n',
+    // `usage` is a sibling of `delta`, and `delta` carries only the stop fields —
+    // see RawMessageDeltaEvent in @anthropic-ai/sdk. The null cache counters are
+    // what the API sends for an uncached request.
+    'event: message_delta\ndata: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":null,"cache_creation_input_tokens":null,"cache_read_input_tokens":null,"output_tokens":1000000}}\n\n',
   ].join("");
 }
diff --git a/services/gateway/test/usage.test.ts b/services/gateway/test/usage.test.ts
index 6cfed2e9..010ef13d 100644
--- a/services/gateway/test/usage.test.ts
+++ b/services/gateway/test/usage.test.ts
@@ -6,7 +6,7 @@ const MESSAGE_START =
 const CONTENT_DELTA =
   'event: content_block_delta\ndata: {"type":"content_block_delta","delta":{"text":"ok"}}\n\n';
 const MESSAGE_DELTA =
-  'event: message_delta\ndata: {"type":"message_delta","delta":{"usage":{"output_tokens":700}}}\n\n';
+  'event: message_delta\ndata: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"output_tokens":700}}\n\n';
 
 describe("UsageTee anthropic streaming usage", () => {
   it("meters input and cache tokens reported by message_start", async () => {