diff --git a/routstr/payment/cost_calculation.py b/routstr/payment/cost_calculation.py
index f28ec6e3..b4b0ff6b 100644
--- a/routstr/payment/cost_calculation.py
+++ b/routstr/payment/cost_calculation.py
@@ -1,7 +1,7 @@
 import math
 from typing import TYPE_CHECKING
 
-from pydantic.v1 import BaseModel
+from pydantic.v1 import BaseModel, Field
 
 from ..core import get_logger
 from ..core.settings import settings
@@ -37,6 +37,7 @@ class CostData(BaseModel):
     cache_creation_msats: int = 0
     # Actual debit after finalization; None means settlement has not run yet.
     charged_msats: int | None = None
+    upstream_usd: float = Field(default=0.0, exclude=True)
 
 
 class MaxCostData(CostData):
@@ -491,6 +492,7 @@ def _calculate_from_usd_cost(
     """Calculate cost from USD figures, deriving input/output split from tokens."""
     if provider_fee is None:
         provider_fee = _resolve_provider_fee(response_data.get("model", ""))
+    reported_usd = usd_cost
     usd_cost = usd_cost * provider_fee
     input_usd = input_usd * provider_fee
     output_usd = output_usd * provider_fee
@@ -576,6 +578,7 @@ def _calculate_from_usd_cost(
         cache_creation_input_tokens=cache_creation_tokens,
         cache_read_msats=cache_read_msats,
         cache_creation_msats=cache_creation_msats,
+        upstream_usd=reported_usd,
     )
 
 
diff --git a/tests/unit/test_upstream_reported_cost.py b/tests/unit/test_upstream_reported_cost.py
new file mode 100644
index 00000000..0ee7f9d0
--- /dev/null
+++ b/tests/unit/test_upstream_reported_cost.py
@@ -0,0 +1,123 @@
+"""Tests that the upstream's own USD figure survives the provider-fee multiply.
+
+``_calculate_from_usd_cost`` multiplies the fee into the same local that holds
+the upstream's reported cost, so by the time a ``CostData`` exists the raw
+figure is gone and only the marked-up one remains. Dividing the total back out
+does not recover it: when the request falls through to token pricing the total
+is the node's own arithmetic, and dividing it compares that number to itself.
+
+These tests cover ``upstream_usd`` — the pre-fee figure carried alongside the
+billed one, and the discriminator for whether an upstream reported a cost at
+all. They also cover the boundary it must not cross: the pair of numbers spells
+out the node's margin, so it stays internal and is never serialised to a client.
+"""
+
+from __future__ import annotations
+
+import math
+from collections.abc import Iterator
+from typing import Any
+from unittest.mock import patch
+
+import pytest
+
+from routstr.payment.cost_calculation import CostData, calculate_cost
+from routstr.payment.models import Architecture, Model, Pricing
+
+
+@pytest.fixture(autouse=True)
+def patch_sats_usd_price() -> Iterator[None]:
+    """Pin the exchange rate; these tests are about the USD figure, not the feed."""
+    with patch("routstr.payment.cost_calculation.sats_usd_price", return_value=5.0e-5):
+        yield
+
+
+def _model() -> Model:
+    return Model(
+        id="m",
+        name="m",
+        created=0,
+        description="d",
+        context_length=8192,
+        architecture=Architecture(
+            modality="text",
+            input_modalities=["text"],
+            output_modalities=["text"],
+            tokenizer="unknown",
+            instruct_type=None,
+        ),
+        pricing=Pricing(prompt=1e-06, completion=2e-06),
+        sats_pricing=Pricing(prompt=1e-06, completion=2e-06),
+    )
+
+
+def _response(usage: dict[str, Any]) -> dict[str, Any]:
+    return {"model": "m", "usage": usage}
+
+
+@pytest.mark.asyncio
+async def test_reported_cost_is_kept_alongside_the_billed_one() -> None:
+    """The billed total carries the fee; ``upstream_usd`` must not."""
+    response = _response(
+        {"prompt_tokens": 1000, "completion_tokens": 500, "cost": 0.01}
+    )
+
+    cost = await calculate_cost(
+        response, max_cost=999999, model_obj=_model(), provider_fee=1.05
+    )
+
+    assert isinstance(cost, CostData)
+    assert cost.total_usd == pytest.approx(0.0105)
+    assert cost.upstream_usd == pytest.approx(0.01)
+
+
+@pytest.mark.asyncio
+async def test_token_priced_request_reports_no_upstream_cost() -> None:
+    """Nothing was reported, so there is nothing to carry — not our own total."""
+    response = _response({"prompt_tokens": 1000, "completion_tokens": 500})
+
+    cost = await calculate_cost(
+        response, max_cost=999999, model_obj=_model(), provider_fee=1.05
+    )
+
+    assert isinstance(cost, CostData)
+    assert cost.total_msats > 0
+    assert cost.upstream_usd == 0.0
+
+
+@pytest.mark.asyncio
+async def test_reported_cost_is_never_serialised_to_a_client() -> None:
+    """Publishing it beside the billed total would spell out the node's margin."""
+    response = _response(
+        {"prompt_tokens": 1000, "completion_tokens": 500, "cost": 0.01}
+    )
+
+    cost = await calculate_cost(
+        response, max_cost=999999, model_obj=_model(), provider_fee=1.05
+    )
+
+    assert isinstance(cost, CostData)
+    assert cost.upstream_usd == pytest.approx(0.01)
+    assert "upstream_usd" not in cost.dict()
+    assert "upstream_usd" not in cost.json()
+
+
+@pytest.mark.asyncio
+async def test_billed_total_is_reproducible_from_the_reported_cost() -> None:
+    """Fee, rate and rounding must carry the reported figure to the billed one.
+
+    The identity a report can restate: whatever the upstream said, times the
+    provider fee, converted at the current rate and rounded up, is what the node
+    charged. A figure chosen for its awkward remainder keeps the ceiling honest.
+    """
+    response = _response(
+        {"prompt_tokens": 1000, "completion_tokens": 500, "cost": 0.000123}
+    )
+
+    cost = await calculate_cost(
+        response, max_cost=999999, model_obj=_model(), provider_fee=1.03
+    )
+
+    assert isinstance(cost, CostData)
+    assert cost.total_msats == math.ceil(cost.upstream_usd * 1.03 / 5.0e-5 * 1000)
+    assert cost.total_msats == 2534