diff --git a/open-sse/utils/usageTracking.ts b/open-sse/utils/usageTracking.ts
index 8e3e969c4e..f1758b7030 100644
--- a/open-sse/utils/usageTracking.ts
+++ b/open-sse/utils/usageTracking.ts
@@ -402,6 +402,18 @@ export function extractUsage(chunk) {
     });
   }
 
+  // Ollama NDJSON format (raw from provider, before translation)
+  // Ollama sends: { "model": "...", "done": true, "prompt_eval_count": N, "eval_count": M }
+  if (chunk.done === true && typeof chunk.prompt_eval_count === "number") {
+    const promptEvalCount = chunk.prompt_eval_count || 0;
+    const evalCount = chunk.eval_count || 0;
+    return normalizeUsage({
+      prompt_tokens: promptEvalCount,
+      completion_tokens: evalCount,
+      total_tokens: promptEvalCount + evalCount,
+    });
+  }
+
   return null;
 }
 
diff --git a/tests/unit/usage-extractor.test.ts b/tests/unit/usage-extractor.test.ts
index 6c7bdf5273..e156a989af 100644
--- a/tests/unit/usage-extractor.test.ts
+++ b/tests/unit/usage-extractor.test.ts
@@ -332,3 +332,45 @@ test("extractUsage reads flat cached_tokens and reasoning_tokens from streaming
   assert.equal(usage.cached_tokens, 192);
   assert.equal(usage.reasoning_tokens, 49);
 });
+
+// ── Ollama raw NDJSON streaming usage ──
+// Ollama sends a final NDJSON line { done: true, prompt_eval_count, eval_count }
+// (raw from the provider, before any OpenAI translation). Without a dedicated
+// branch, extractUsage returns null and Ollama streaming usage is dropped.
+
+test("extractUsage reads Ollama raw NDJSON final chunk (done + prompt_eval_count/eval_count)", () => {
+  const usage = extractUsage({
+    model: "llama3.1",
+    done: true,
+    prompt_eval_count: 26,
+    eval_count: 298,
+  });
+
+  assert.ok(usage, "expected usage to be extracted from the Ollama final chunk");
+  assert.equal(usage.prompt_tokens, 26);
+  assert.equal(usage.completion_tokens, 298);
+  assert.equal(usage.total_tokens, 324);
+});
+
+test("extractUsage defaults missing Ollama eval counts to zero", () => {
+  const usage = extractUsage({
+    model: "llama3.1",
+    done: true,
+    prompt_eval_count: 12,
+  });
+
+  assert.ok(usage, "expected usage to be extracted even with only prompt_eval_count");
+  assert.equal(usage.prompt_tokens, 12);
+  assert.equal(usage.completion_tokens, 0);
+  assert.equal(usage.total_tokens, 12);
+});
+
+test("extractUsage ignores non-final Ollama NDJSON chunks (done=false)", () => {
+  const usage = extractUsage({
+    model: "llama3.1",
+    done: false,
+    response: "partial",
+  });
+
+  assert.equal(usage, null);
+});