Merge branch 'fix-too-low-request-costs' into token-storage-and-rate-limit

routstr/payment/cost_calculation.py

@@ -374,16 +374,38 @@ def _calculate_from_usd_cost(
         input_msats = math.floor(cost_in_msats * input_usd / component_usd)
         output_msats = cost_in_msats - input_msats
     else:
-        effective_input_tokens = (
-            input_tokens + cache_read_tokens + cache_creation_tokens
+        # Providers often report only a total USD cost. Derive the visible
+        # input/output split from the model's relative token prices; raw token
+        # counts alone are misleading when completion tokens cost more.
+        try:
+            pricing_rates = _get_pricing_rates(response_data)
+        except ValueError:
+            pricing_rates = None
+
+        if pricing_rates is None:
+            input_rate = float(settings.fixed_per_1k_input_tokens) * 1000.0
+            output_rate = float(settings.fixed_per_1k_output_tokens) * 1000.0
+            cache_read_rate = input_rate
+            cache_creation_rate = input_rate
+        else:
+            input_rate, output_rate, cache_read_rate, cache_creation_rate = (
+                pricing_rates
+            )
+
+        input_weight = (
+            input_tokens * input_rate
+            + cache_read_tokens * cache_read_rate
+            + cache_creation_tokens * cache_creation_rate
         )
-        total_tokens = effective_input_tokens + output_tokens
-        input_msats = (
-            int(cost_in_msats * effective_input_tokens / total_tokens)
-            if total_tokens > 0
-            else 0
-        )
-        output_msats = cost_in_msats - input_msats
+        output_weight = output_tokens * output_rate
+        total_weight = input_weight + output_weight
+
+        if total_weight > 0:
+            input_msats = math.floor(cost_in_msats * input_weight / total_weight)
+            output_msats = cost_in_msats - input_msats
+        else:
+            input_msats = 0
+            output_msats = cost_in_msats
 
     logger.info(
         "Using cost from usage data/details",

tests/unit/test_cost_calculation_caching.py

@@ -5,7 +5,7 @@ edge cases, and billing accuracy.
 """
 
 import os
-from unittest.mock import patch
+from unittest.mock import Mock, patch
 
 import pytest
 
@@ -500,6 +500,38 @@ async def test_small_usd_cost_components_sum_to_rounded_total(
     assert result.input_msats + result.output_msats == result.total_msats
 
 
+@pytest.mark.asyncio
+async def test_total_only_usd_cost_uses_model_prices_for_component_split(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """A reported total is split by priced tokens, not raw token counts."""
+    monkeypatch.setattr(settings, "fixed_pricing", False)
+    response = {
+        "model": "z-ai/glm-5.2-20260616",
+        "usage": {
+            "prompt_tokens": 375,
+            "completion_tokens": 218,
+            "total_cost": 0.00039088,
+        },
+    }
+    pricing = Mock(
+        prompt=0.0001,
+        completion=0.001,
+        input_cache_read=0.0001,
+        input_cache_write=0.0001,
+    )
+    model = Mock(sats_pricing=pricing)
+
+    with patch("routstr.proxy.get_model_instance", return_value=model):
+        result = await calculate_cost(response, max_cost=100000)
+
+    assert isinstance(result, CostData)
+    assert result.total_msats == 7818
+    assert result.input_msats + result.output_msats == result.total_msats
+    assert result.input_msats == 1147
+    assert result.output_msats == 6671
+
+
 # ============================================================================
 # Test 13: Missing Usage Block
 # ============================================================================