Add cached input pricing support in cost calculations. Closes #1582

Author: waldekmastykarz

DevProxy.Abstractions/LanguageModel/PricesData.cs

@@ -9,6 +9,7 @@ namespace DevProxy.Abstractions.LanguageModel;
 public class ModelPrices
 {
     public double Input { get; set; }
+    public double CachedInput { get; set; }
     public double Output { get; set; }
 }
 
@@ -44,7 +45,7 @@ public bool TryGetModelPrices(string modelName, out ModelPrices? prices)
         return false;
     }
 
-    public (double Input, double Output) CalculateCost(string modelName, long inputTokens, long outputTokens)
+    public (double Input, double Output) CalculateCost(string modelName, long inputTokens, long outputTokens, long cachedInputTokens = 0)
     {
         if (!TryGetModelPrices(modelName, out var prices))
         {
@@ -54,7 +55,14 @@ public bool TryGetModelPrices(string modelName, out ModelPrices? prices)
         Debug.Assert(prices != null, "Prices data should not be null here.");
 
         // Prices in the data are per 1M tokens
-        var inputCost = prices.Input * (inputTokens / 1_000_000.0);
+        // When cached input pricing is available, separate cached tokens
+        // from regular input tokens for accurate cost calculation
+        var regularInputTokens = inputTokens - cachedInputTokens;
+        var inputCost = prices.Input * (regularInputTokens / 1_000_000.0);
+        if (cachedInputTokens > 0 && prices.CachedInput > 0)
+        {
+            inputCost += prices.CachedInput * (cachedInputTokens / 1_000_000.0);
+        }
         var outputCost = prices.Output * (outputTokens / 1_000_000.0);
 
         return (inputCost, outputCost);

DevProxy.Plugins/Inspection/LanguageModelPricingLoader.cs

@@ -40,11 +40,18 @@ protected override void LoadData(string fileContents)
                     if (modelProperty.Value.TryGetProperty("input", out var inputElement) &&
                         modelProperty.Value.TryGetProperty("output", out var outputElement))
                     {
-                        pricesData[modelName] = new()
+                        var modelPrices = new ModelPrices
                         {
                             Input = inputElement.GetDouble(),
                             Output = outputElement.GetDouble()
                         };
+
+                        if (modelProperty.Value.TryGetProperty("cached_input", out var cachedInputElement))
+                        {
+                            modelPrices.CachedInput = cachedInputElement.GetDouble();
+                        }
+
+                        pricesData[modelName] = modelPrices;
                     }
                 }
 

DevProxy.Plugins/Inspection/OpenAITelemetryPlugin.cs

@@ -931,7 +931,8 @@ private void RecordUsageMetrics(Activity activity, OpenAIRequest request, OpenAI
             return;
         }
 
-        var (inputCost, outputCost) = Configuration.Prices.CalculateCost(response.Model, usage.PromptTokens, usage.CompletionTokens);
+        var cachedTokens = usage.PromptTokensDetails?.CachedTokens ?? 0L;
+        var (inputCost, outputCost) = Configuration.Prices.CalculateCost(response.Model, usage.PromptTokens, usage.CompletionTokens, cachedTokens);
 
         if (inputCost > 0)
         {
@@ -1042,7 +1043,8 @@ private List<OpenAITelemetryPluginReportModelUsageInformation> GetReportModelUsa
             return usagePerModel;
         }
 
-        var (inputCost, outputCost) = Configuration.Prices.CalculateCost(response.Model, usage.PromptTokens, usage.CompletionTokens);
+        var cachedTokens = usage.PromptTokensDetails?.CachedTokens ?? 0L;
+        var (inputCost, outputCost) = Configuration.Prices.CalculateCost(response.Model, usage.PromptTokens, usage.CompletionTokens, cachedTokens);
 
         if (inputCost > 0)
         {