diff --git a/lib/server/proxy/responses.ts b/lib/server/proxy/responses.ts
index 07b5292..b82583e 100644
--- a/lib/server/proxy/responses.ts
+++ b/lib/server/proxy/responses.ts
@@ -1046,6 +1046,83 @@ const normalizeTranscriptMessageToolNames = (
   });
 };
 
+const toResponsesUsageNumber = (value: unknown): number => {
+  const numeric =
+    typeof value === 'number' ? value : Number.parseFloat(String(value ?? ''));
+
+  if (!Number.isFinite(numeric) || numeric < 0) {
+    return 0;
+  }
+
+  return numeric;
+};
+
+const mapChatUsageToResponses = (usage: unknown): Record<string, unknown> => {
+  if (!usage || typeof usage !== 'object') {
+    return {
+      input_tokens: 0,
+      input_tokens_details: { cached_tokens: 0 },
+      output_tokens: 0,
+      output_tokens_details: { reasoning_tokens: 0 },
+      total_tokens: 0,
+    };
+  }
+
+  const value = usage as {
+    cache_creation_input_tokens?: unknown;
+    cache_read_input_tokens?: unknown;
+    completion_tokens?: unknown;
+    completion_tokens_details?: { reasoning_tokens?: unknown };
+    completion_thinking_tokens?: unknown;
+    input_tokens_details?: { cached_tokens?: unknown };
+    prompt_cache_hit_tokens?: unknown;
+    prompt_cache_miss_tokens?: unknown;
+    prompt_cache_write_tokens?: unknown;
+    prompt_tokens?: unknown;
+    prompt_tokens_details?: {
+      cache_creation_tokens?: unknown;
+      cached_tokens?: unknown;
+    };
+    total_tokens?: unknown;
+  };
+  const outputTokens = toResponsesUsageNumber(value.completion_tokens);
+  const cachedTokens = toResponsesUsageNumber(
+    value.prompt_tokens_details?.cached_tokens ??
+      value.input_tokens_details?.cached_tokens ??
+      value.cache_read_input_tokens ??
+      value.prompt_cache_hit_tokens,
+  );
+  const cacheCreationTokens = toResponsesUsageNumber(
+    value.prompt_tokens_details?.cache_creation_tokens ??
+      value.cache_creation_input_tokens ??
+      value.prompt_cache_write_tokens,
+  );
+  const reasoningTokens = toResponsesUsageNumber(
+    value.completion_tokens_details?.reasoning_tokens ??
+      value.completion_thinking_tokens,
+  );
+  // Chat usage is the single source of truth for both shapes. Keep the
+  // Responses counters faithful to it so clients never see zeroed metrics.
+  // prompt_tokens already covers its cached and created subsets, so the
+  // split counters are only summed when prompt_tokens is missing. Otherwise
+  // cached tokens would exceed the reported input total.
+  const inputTokens = toResponsesUsageNumber(
+    value.prompt_tokens ??
+      toResponsesUsageNumber(value.prompt_cache_miss_tokens) +
+        cachedTokens +
+        cacheCreationTokens,
+  );
+
+  return {
+    input_tokens: inputTokens,
+    input_tokens_details: { cached_tokens: cachedTokens },
+    output_tokens: outputTokens,
+    output_tokens_details: { reasoning_tokens: reasoningTokens },
+    total_tokens:
+      toResponsesUsageNumber(value.total_tokens) || inputTokens + outputTokens,
+  };
+};
+
 const mapChatResponseToResponsesPayload = async (
   accessKeyId: string | null,
   credentialFilename: string | null,
@@ -1127,7 +1204,7 @@ const mapChatResponseToResponsesPayload = async (
     model,
     output,
     output_text: outputText,
-    usage: upstreamPayload.usage ?? null,
+    usage: mapChatUsageToResponses(upstreamPayload.usage),
     metadata: defaults.metadata ?? {},
     previous_response_id: previousResponseId,
   };
@@ -1180,6 +1257,7 @@ const createResponsesEventStream = async (
   const toolCallStates = new Map<string, StreamingToolCallState>();
   const toolCallStateKeys = new Map<string, string>();
   let nextToolCallOutputIndex = 1;
+  let latestUsage: unknown = null;
   let reader: ReadableStreamDefaultReader<Uint8Array> | null = null;
   let cancelled = false;
   const releaseReader = (): void => {
@@ -1405,6 +1483,7 @@ const createResponsesEventStream = async (
                 status: 'completed',
                 output_text: outputText,
                 previous_response_id: previousResponseId,
+                usage: mapChatUsageToResponses(latestUsage),
                 output: [
                   ...(outputText
                     ? [buildStreamingMessageItem('completed')]
@@ -1464,7 +1543,13 @@ const createResponsesEventStream = async (
                     tool_calls?: ChatResponseToolCall[];
                   };
                 }>;
+                usage?: unknown;
               };
+              // The final upstream chunk carries the aggregated usage, so
+              // remember it for the downstream response.completed event.
+              if (payload.usage !== undefined) {
+                latestUsage = payload.usage;
+              }
               const delta = payload.choices?.[0]?.delta;
 
               if (delta?.content) {
diff --git a/tests/server/responses-memory.test.ts b/tests/server/responses-memory.test.ts
index 32b4487..851be01 100644
--- a/tests/server/responses-memory.test.ts
+++ b/tests/server/responses-memory.test.ts
@@ -339,6 +339,148 @@ describe('Responses memory bounds', () => {
     expect(writePgSession).not.toHaveBeenCalled();
   });
 
+  it('maps Chat usage onto Responses usage for non-streamed requests', async () => {
+    vi.spyOn(globalThis, 'fetch').mockResolvedValueOnce(
+      new Response(
+        JSON.stringify({
+          choices: [{ message: { content: 'usage answer' } }],
+          usage: {
+            completion_tokens: 506,
+            completion_tokens_details: { reasoning_tokens: 12 },
+            prompt_tokens: 281734,
+            prompt_tokens_details: { cached_tokens: 281408 },
+            total_tokens: 282240,
+          },
+        }),
+        { headers: { 'Content-Type': 'application/json' } },
+      ),
+    );
+
+    const response = await handleResponsesRequest(makeRequest(), {
+      input: 'usage please',
+      model: 'gpt-5.5',
+    });
+    const payload = (await response.json()) as Record<string, unknown>;
+
+    expect(payload.usage).toEqual({
+      input_tokens: 281734,
+      input_tokens_details: { cached_tokens: 281408 },
+      output_tokens: 506,
+      output_tokens_details: { reasoning_tokens: 12 },
+      total_tokens: 282240,
+    });
+  });
+
+  it('emits mapped usage in streamed response.completed events', async () => {
+    vi.spyOn(globalThis, 'fetch').mockResolvedValueOnce(
+      new Response(
+        'data: {"choices":[{"delta":{"content":"usage "}}]}\n\n' +
+          'data: {"choices":[{"delta":{"content":"stream"}}]}\n\n' +
+          'data: {"choices":[],"usage":{"prompt_tokens":281734,"completion_tokens":506,"total_tokens":282240,"prompt_tokens_details":{"cached_tokens":281408},"prompt_cache_hit_tokens":281408,"prompt_cache_miss_tokens":326,"completion_tokens_details":{"reasoning_tokens":12}}}\n\n' +
+          'data: [DONE]\n\n',
+        { headers: { 'Content-Type': 'text/event-stream' } },
+      ),
+    );
+
+    const response = await handleResponsesRequest(makeRequest(), {
+      input: 'stream usage please',
+      model: 'gpt-5.5',
+      stream: true,
+    });
+    const body = await response.text();
+
+    expect(body).toContain('"output_text":"usage stream"');
+    expect(body).toContain('"input_tokens":281734');
+    expect(body).toContain('"cached_tokens":281408');
+    expect(body).toContain('"output_tokens":506');
+    expect(body).toContain('"reasoning_tokens":12');
+    expect(body).toContain('"total_tokens":282240');
+  });
+
+  it('reports zeroed Responses usage when upstream omits usage', async () => {
+    vi.spyOn(globalThis, 'fetch').mockResolvedValueOnce(
+      makeChatResponse('no usage here'),
+    );
+
+    const response = await handleResponsesRequest(makeRequest(), {
+      input: 'missing usage',
+      model: 'gpt-5.5',
+    });
+    const payload = (await response.json()) as Record<string, unknown>;
+
+    expect(payload.usage).toEqual({
+      input_tokens: 0,
+      input_tokens_details: { cached_tokens: 0 },
+      output_tokens: 0,
+      output_tokens_details: { reasoning_tokens: 0 },
+      total_tokens: 0,
+    });
+  });
+
+  it('does not double-count cache creation in the fallback total', async () => {
+    vi.spyOn(globalThis, 'fetch').mockResolvedValueOnce(
+      new Response(
+        JSON.stringify({
+          choices: [{ message: { content: 'derived totals' } }],
+          usage: {
+            completion_tokens: 3,
+            prompt_tokens: 10,
+            prompt_tokens_details: { cache_creation_tokens: 2 },
+          },
+        }),
+        { headers: { 'Content-Type': 'application/json' } },
+      ),
+    );
+
+    const response = await handleResponsesRequest(makeRequest(), {
+      input: 'derive totals',
+      model: 'gpt-5.5',
+    });
+    const payload = (await response.json()) as Record<string, unknown>;
+
+    // prompt_tokens already includes cache creation, so it must not be
+    // added again when computing the fallback total.
+    expect(payload.usage).toEqual({
+      input_tokens: 10,
+      input_tokens_details: { cached_tokens: 0 },
+      output_tokens: 3,
+      output_tokens_details: { reasoning_tokens: 0 },
+      total_tokens: 13,
+    });
+  });
+
+  it('sums split cache counters when upstream omits prompt_tokens', async () => {
+    vi.spyOn(globalThis, 'fetch').mockResolvedValueOnce(
+      new Response(
+        JSON.stringify({
+          choices: [{ message: { content: 'split counters' } }],
+          usage: {
+            completion_tokens: 506,
+            prompt_cache_hit_tokens: 281408,
+            prompt_cache_miss_tokens: 326,
+          },
+        }),
+        { headers: { 'Content-Type': 'application/json' } },
+      ),
+    );
+
+    const response = await handleResponsesRequest(makeRequest(), {
+      input: 'split counters',
+      model: 'gpt-5.5',
+    });
+    const payload = (await response.json()) as Record<string, unknown>;
+
+    // Without prompt_tokens, the split counters must be summed so cached
+    // tokens never exceed the reported input total.
+    expect(payload.usage).toEqual({
+      input_tokens: 281734,
+      input_tokens_details: { cached_tokens: 281408 },
+      output_tokens: 506,
+      output_tokens_details: { reasoning_tokens: 0 },
+      total_tokens: 282240,
+    });
+  });
+
   it('bounds incomplete SSE frames in every proxy stream', async () => {
     const oversizedFrame = 'x'.repeat(1_000_001);
     const fetchMock = vi