fix: improve cost tracking accuracy for modern models

- Added support for reasoning tokens in cost calculations.
- Fixed DeepSeek cache-write token mapping (PromptCacheMissTokens).
- Improved CalculateCost debug logging to trace all pricing variables.

internal/providers/deepseek.go

@@ -51,6 +51,9 @@ func (u *deepSeekUsage) ToUnified() *models.Usage {
 	if u.PromptCacheHitTokens > 0 {
 		usage.CacheReadTokens = &u.PromptCacheHitTokens
 	}
+	if u.PromptCacheMissTokens > 0 {
+		usage.CacheWriteTokens = &u.PromptCacheMissTokens
+	}
 	if u.CompletionTokensDetails != nil && u.CompletionTokensDetails.ReasoningTokens > 0 {
 		usage.ReasoningTokens = &u.CompletionTokensDetails.ReasoningTokens
 	}

internal/server/server.go

@@ -378,9 +378,9 @@ func (s *Server) logRequest(start time.Time, clientID, provider, model string, u
 		}
 		
 		// Calculate cost using registry
-		entry.Cost = utils.CalculateCost(s.registry, model, entry.PromptTokens, entry.CompletionTokens, entry.CacheReadTokens, entry.CacheWriteTokens)
-		fmt.Printf("[DEBUG] Request logged: model=%s, prompt=%d, completion=%d, cache_read=%d, cost=%f\n", 
-			model, entry.PromptTokens, entry.CompletionTokens, entry.CacheReadTokens, entry.Cost)
+		entry.Cost = utils.CalculateCost(s.registry, model, entry.PromptTokens, entry.CompletionTokens, entry.ReasoningTokens, entry.CacheReadTokens, entry.CacheWriteTokens)
+		fmt.Printf("[DEBUG] Request logged: model=%s, prompt=%d, completion=%d, reasoning=%d, cache_read=%d, cost=%f\n", 
+			model, entry.PromptTokens, entry.CompletionTokens, entry.ReasoningTokens, entry.CacheReadTokens, entry.Cost)
 	}
 
 	s.logger.LogRequest(entry)

internal/utils/registry.go

@@ -34,9 +34,10 @@ func FetchRegistry() (*models.ModelRegistry, error) {
 	return &models.ModelRegistry{Providers: providers}, nil
 }
 
-func CalculateCost(registry *models.ModelRegistry, modelID string, promptTokens, completionTokens, cacheRead, cacheWrite uint32) float64 {
+func CalculateCost(registry *models.ModelRegistry, modelID string, promptTokens, completionTokens, reasoningTokens, cacheRead, cacheWrite uint32) float64 {
 	meta := registry.FindModel(modelID)
 	if meta == nil || meta.Cost == nil {
+		log.Printf("[DEBUG] CalculateCost: model %s not found or has no cost metadata", modelID)
 		return 0.0
 	}
 
@@ -61,5 +62,8 @@ func CalculateCost(registry *models.ModelRegistry, modelID string, promptTokens,
 		cost += float64(cacheWrite) * (*meta.Cost.CacheWrite) / 1000000.0
 	}
 
+	log.Printf("[DEBUG] CalculateCost: model=%s, uncached=%d, completion=%d, reasoning=%d, cache_read=%d, cache_write=%d, cost=%f (input_rate=%f, output_rate=%f)", 
+		modelID, uncachedTokens, completionTokens, reasoningTokens, cacheRead, cacheWrite, cost, meta.Cost.Input, meta.Cost.Output)
+
 	return cost
 }