fix

apps/api/internal/processing/ai.go

@@ -43,10 +43,14 @@ type EnableChecker interface {
 type Completion struct {
 	Text         string
 	PromptTokens int
-	OutputTokens int
-	TotalTokens  int
-	Model        string
-	Raw          any
+	// CachedPromptTokens is the cached subset of PromptTokens (OpenAI
+	// usage.prompt_tokens_details.cached_tokens). Billed far cheaper than fresh
+	// input, so cost accounting must subtract it. 0 when the provider omits it.
+	CachedPromptTokens int
+	OutputTokens       int
+	TotalTokens        int
+	Model              string
+	Raw                any
 }
 
 // Embedder turns text into vectors (OpenAI-compatible /v1/embeddings).

apps/api/internal/processing/openai.go

@@ -242,6 +242,11 @@ type chatResponse struct {
 		PromptTokens     int `json:"prompt_tokens"`
 		CompletionTokens int `json:"completion_tokens"`
 		TotalTokens      int `json:"total_tokens"`
+		// Cached prompt tokens bill at roughly a tenth of fresh input, and enhance
+		// re-sends a large shared system prompt, so cost is wrong without them.
+		PromptTokensDetails struct {
+			CachedTokens int `json:"cached_tokens"`
+		} `json:"prompt_tokens_details"`
 	} `json:"usage"`
 	Error *openAIErrorBody `json:"error"`
 }
@@ -557,11 +562,12 @@ func (c *OpenAIClient) doComplete(ctx context.Context, system, user string, temp
 	log.Printf("openai: chat content ok model=%s finish_reason=%s content_runes=%d prompt_tokens=%d completion_tokens=%d total_tokens=%d",
 		c.Model, finishReason, len([]rune(text)), usagePrompt, usageOut, usageTotal)
 	return Completion{
-		Text:         text,
-		PromptTokens: usagePrompt,
-		OutputTokens: usageOut,
-		TotalTokens:  usageTotal,
-		Model:        parsed.Model,
+		Text:               text,
+		PromptTokens:       usagePrompt,
+		CachedPromptTokens: parsed.Usage.PromptTokensDetails.CachedTokens,
+		OutputTokens:       usageOut,
+		TotalTokens:        usageTotal,
+		Model:              parsed.Model,
 		Raw: map[string]any{
 			"model":         parsed.Model,
 			"usage":         parsed.Usage,