fix(proxy): populate contribution USD savings

headroom/proxy/handlers/anthropic.py

@@ -2379,6 +2379,10 @@ async def api_call_fn(
                     if _auth_header.startswith("Bearer ") and not _auth_header.startswith(
                         "Bearer sk-ant-api"
                     ):
+                        from headroom.proxy.savings_tracker import (
+                            _estimate_cache_savings_usd,
+                            _estimate_compression_savings_usd,
+                        )
                         from headroom.subscription.tracker import (
                             get_subscription_tracker as _get_sub_tracker,
                         )
@@ -2389,6 +2393,10 @@ async def api_call_fn(
                                 tokens_submitted=optimized_tokens,
                                 tokens_saved_compression=tokens_saved,
                                 tokens_saved_cache_reads=cr_tokens,
+                                compression_savings_usd=_estimate_compression_savings_usd(
+                                    model, tokens_saved
+                                ),
+                                cache_savings_usd=_estimate_cache_savings_usd(model, cr_tokens),
                             )
 
                     # The pre-refactor PERF emit (above) read raw usage

headroom/proxy/savings_tracker.py

@@ -202,6 +202,28 @@ def _estimate_compression_savings_usd(model: str, tokens_saved: int) -> float:
         return 0.0
 
 
+def _estimate_cache_savings_usd(model: str, cache_read_tokens: int) -> float:
+    """Estimate cache-read savings in USD from discounted input tokens."""
+
+    litellm = _get_litellm_module()
+    if cache_read_tokens <= 0 or litellm is None:
+        return 0.0
+
+    try:
+        resolved = _resolve_litellm_model(model)
+        info = litellm.model_cost.get(resolved, {})
+        input_cost_per_token = info.get("input_cost_per_token")
+        if not input_cost_per_token:
+            return 0.0
+        cache_read_cost_per_token = info.get("cache_read_input_token_cost", input_cost_per_token)
+        savings_per_token = float(input_cost_per_token) - float(cache_read_cost_per_token)
+        if savings_per_token <= 0:
+            return 0.0
+        return float(cache_read_tokens) * savings_per_token
+    except Exception:
+        return 0.0
+
+
 def _estimate_input_cost_usd(
     model: str,
     input_tokens: int,

tests/test_proxy_cache_savings_usd.py

@@ -0,0 +1,35 @@
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import pytest
+
+from headroom.proxy import savings_tracker as savings_tracker_module
+
+
+def test_estimate_cache_savings_usd_uses_discounted_cache_read_price(monkeypatch) -> None:
+    fake_litellm = SimpleNamespace(
+        model_cost={
+            "gpt-4o": {
+                "input_cost_per_token": 0.002,
+                "cache_read_input_token_cost": 0.001,
+            }
+        }
+    )
+    monkeypatch.setattr(savings_tracker_module, "LITELLM_AVAILABLE", True)
+    monkeypatch.setattr(savings_tracker_module, "litellm", fake_litellm)
+
+    assert savings_tracker_module._estimate_cache_savings_usd("gpt-4o", 100) == pytest.approx(
+        0.1
+    )
+
+
+def test_estimate_cache_savings_usd_handles_missing_pricing(monkeypatch) -> None:
+    fake_litellm = SimpleNamespace(model_cost={})
+    monkeypatch.setattr(savings_tracker_module, "LITELLM_AVAILABLE", True)
+    monkeypatch.setattr(savings_tracker_module, "litellm", fake_litellm)
+
+    assert savings_tracker_module._estimate_cache_savings_usd("gpt-4o", 100) == 0.0
+
+    monkeypatch.setattr(savings_tracker_module, "LITELLM_AVAILABLE", False)
+    assert savings_tracker_module._estimate_cache_savings_usd("gpt-4o", 100) == 0.0