diff --git a/app/routes/chat.py b/app/routes/chat.py
index 1779d44..26236d8 100644
--- a/app/routes/chat.py
+++ b/app/routes/chat.py
@@ -148,8 +148,8 @@ async def _build_log_row(
     meta = response.get("_orca_meta", {}) or {}
     usage = response.get("usage", {}) or {}
     resolved = actual_resolved or response.get("model") or requested_model
-    input_t = usage.get("prompt_tokens", 0) or 0
-    output_t = usage.get("completion_tokens", 0) or 0
+    input_t = max(0, usage.get("prompt_tokens", 0) or 0)
+    output_t = max(0, usage.get("completion_tokens", 0) or 0)
     return RequestLog(
         workspace_id=str(kc.workspace_id),
         api_key_id=str(kc.key_id),
diff --git a/tests/unit/test_negative_token_clamping.py b/tests/unit/test_negative_token_clamping.py
new file mode 100644
index 0000000..d76fc69
--- /dev/null
+++ b/tests/unit/test_negative_token_clamping.py
@@ -0,0 +1,88 @@
+"""Regression: _build_log_row must clamp negative token counts to zero.
+
+Some upstream providers return negative token counts in malformed usage
+data. Without clamping, the RequestLog row persists negative values which
+corrupts analytics dashboards (negative total tokens, negative spend
+aggregates). The cost calculation already clamps via _compute_cost_microcents,
+but the raw token count columns were unprotected."""
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from app.routes.chat import _build_log_row
+from app.schemas import ChatCompletionRequest, ChatMessage
+from packages.auth.types import KeyContext
+
+
+def _make_body(**overrides):
+    defaults = {
+        "model": "gpt-4o",
+        "messages": [ChatMessage(role="user", content="hello")],
+    }
+    defaults.update(overrides)
+    return ChatCompletionRequest(**defaults)
+
+
+def _make_kc():
+    kc = MagicMock(spec=KeyContext)
+    kc.workspace_id = "ws-1"
+    kc.key_id = "key-1"
+    return kc
+
+
+@pytest.mark.asyncio
+async def test_build_log_row_clamps_negative_input_tokens():
+    body = _make_body()
+    kc = _make_kc()
+    response = {
+        "model": "gpt-4o",
+        "usage": {"prompt_tokens": -5, "completion_tokens": 10},
+        "_orca_meta": {"provider": "openai", "latency_ms": 100},
+    }
+    log = await _build_log_row(
+        body=body, kc=kc, response=response,
+        status_code=200, error_type=None,
+        started_perf=0, strategy="balanced",
+        requested_model="gpt-4o",
+    )
+    assert log.input_tokens == 0
+    assert log.output_tokens == 10
+
+
+@pytest.mark.asyncio
+async def test_build_log_row_clamps_negative_output_tokens():
+    body = _make_body()
+    kc = _make_kc()
+    response = {
+        "model": "gpt-4o",
+        "usage": {"prompt_tokens": 10, "completion_tokens": -3},
+        "_orca_meta": {"provider": "openai", "latency_ms": 100},
+    }
+    log = await _build_log_row(
+        body=body, kc=kc, response=response,
+        status_code=200, error_type=None,
+        started_perf=0, strategy="balanced",
+        requested_model="gpt-4o",
+    )
+    assert log.input_tokens == 10
+    assert log.output_tokens == 0
+
+
+@pytest.mark.asyncio
+async def test_build_log_row_clamps_both_negative():
+    body = _make_body()
+    kc = _make_kc()
+    response = {
+        "model": "gpt-4o",
+        "usage": {"prompt_tokens": -1, "completion_tokens": -1},
+        "_orca_meta": {"provider": "openai", "latency_ms": 100},
+    }
+    log = await _build_log_row(
+        body=body, kc=kc, response=response,
+        status_code=200, error_type=None,
+        started_perf=0, strategy="balanced",
+        requested_model="gpt-4o",
+    )
+    assert log.input_tokens == 0
+    assert log.output_tokens == 0