feat: support cached input token billing

apps/api/internal/runner/limits_test.go

@@ -1,6 +1,7 @@
 package runner
 
 import (
+	"context"
 	"testing"
 
 	"github.com/easyai/easyai-ai-gateway/apps/api/internal/clients"
@@ -29,6 +30,44 @@ func TestTokenUsageAmountsFallsBackToInputOutputTotal(t *testing.T) {
 	}
 }
 
+func TestBillingsSplitsCachedInputTokens(t *testing.T) {
+	service := &Service{}
+	candidate := store.RuntimeModelCandidate{
+		ModelName: "test-model",
+		BillingConfig: map[string]any{
+			"textInputPer1k":       1.0,
+			"textCachedInputPer1k": 0.25,
+			"textOutputPer1k":      2.0,
+		},
+	}
+	response := clients.Response{
+		Usage: clients.Usage{
+			InputTokens:       1000,
+			CachedInputTokens: 400,
+			OutputTokens:      500,
+			TotalTokens:       1500,
+		},
+	}
+
+	lines := service.billings(context.Background(), nil, "chat.completions", nil, candidate, response, false)
+
+	if len(lines) != 3 {
+		t.Fatalf("expected uncached input, cached input, and output lines, got %+v", lines)
+	}
+	uncached, _ := lines[0].(map[string]any)
+	cached, _ := lines[1].(map[string]any)
+	output, _ := lines[2].(map[string]any)
+	if uncached["resourceType"] != "text_input" || uncached["quantity"] != 600 || uncached["amount"] != 0.6 {
+		t.Fatalf("unexpected uncached input billing line: %+v", uncached)
+	}
+	if cached["resourceType"] != "text_cached_input" || cached["quantity"] != 400 || cached["amount"] != 0.1 {
+		t.Fatalf("unexpected cached input billing line: %+v", cached)
+	}
+	if output["resourceType"] != "text_output" || output["quantity"] != 500 || output["amount"] != 1.0 {
+		t.Fatalf("unexpected output billing line: %+v", output)
+	}
+}
+
 func TestEffectiveRateLimitPolicyTreatsEmptyRuntimePolicyAsUnlimited(t *testing.T) {
 	policy := effectiveRateLimitPolicy(store.RuntimeModelCandidate{
 		PlatformRateLimitPolicy: map[string]any{"rules": []any{

apps/api/internal/runner/pricing.go

@@ -62,6 +62,7 @@ func (s *Service) billings(ctx context.Context, user *auth.User, kind string, bo
 	if isTextBillingKind(kind) {
 		inputTokens := response.Usage.InputTokens
 		outputTokens := response.Usage.OutputTokens
+		cachedInputTokens := response.Usage.CachedInputTokens
 		if isTextInputOnlyKind(kind) && inputTokens == 0 && response.Usage.TotalTokens > 0 {
 			inputTokens = response.Usage.TotalTokens
 		}
@@ -71,8 +72,37 @@ func (s *Service) billings(ctx context.Context, user *auth.User, kind string, bo
 				outputTokens = 1
 			}
 		}
-		inputAmount := roundPrice(float64(inputTokens) / 1000 * resourcePrice(config, "text", "textInputPer1k", "inputTokenPrice", "basePrice") * discount)
-		lines := []any{billingLine(candidate, "text_input", "1k_tokens", inputTokens, inputAmount, discount, simulated)}
+		if cachedInputTokens > inputTokens && inputTokens > 0 {
+			cachedInputTokens = inputTokens
+		}
+		uncachedInputTokens := inputTokens - cachedInputTokens
+		if uncachedInputTokens < 0 {
+			uncachedInputTokens = 0
+		}
+		inputPrice := resourcePrice(config, "text", "textInputPer1k", "inputTokenPrice", "basePrice")
+		cachedInputPrice := resourcePrice(config, "text", "textCachedInputPer1k", "cachedInputTokenPrice", "cacheInputTokenPrice", "textInputCacheHitPer1k", "inputCacheHitTokenPrice")
+		if cachedInputPrice <= 0 {
+			cachedInputPrice = inputPrice
+		}
+		inputAmount := roundPrice(float64(uncachedInputTokens) / 1000 * inputPrice * discount)
+		lines := []any{}
+		if uncachedInputTokens > 0 || cachedInputTokens == 0 {
+			lines = append(lines, billingLineWithDetails(candidate, "text_input", "1k_tokens", uncachedInputTokens, inputAmount, discount, simulated, map[string]any{
+				"inputTokens":         inputTokens,
+				"uncachedInputTokens": uncachedInputTokens,
+				"cachedInputTokens":   cachedInputTokens,
+				"pricePer1k":          inputPrice,
+			}))
+		}
+		if cachedInputTokens > 0 {
+			cachedInputAmount := roundPrice(float64(cachedInputTokens) / 1000 * cachedInputPrice * discount)
+			lines = append(lines, billingLineWithDetails(candidate, "text_cached_input", "1k_tokens", cachedInputTokens, cachedInputAmount, discount, simulated, map[string]any{
+				"inputTokens":         inputTokens,
+				"uncachedInputTokens": uncachedInputTokens,
+				"cachedInputTokens":   cachedInputTokens,
+				"pricePer1k":          cachedInputPrice,
+			}))
+		}
 		if isTextGenerationKind(kind) {
 			outputAmount := roundPrice(float64(outputTokens) / 1000 * resourcePrice(config, "text", "textOutputPer1k", "outputTokenPrice", "basePrice") * discount)
 			lines = append(lines, billingLine(candidate, "text_output", "1k_tokens", outputTokens, outputAmount, discount, simulated))
@@ -234,10 +264,22 @@ func resourcePrice(config map[string]any, resource string, keys ...string) float
 				return value
 			}
 		}
+		if formulaConfig, ok := resourceConfig["formulaConfig"].(map[string]any); ok {
+			for _, key := range keys {
+				if value := floatFromAny(formulaConfig[key]); value > 0 {
+					return value
+				}
+			}
+		}
 		if value := floatFromAny(resourceConfig["basePrice"]); value > 0 {
 			return value
 		}
 	}
+	if resource == "text" {
+		if value := resourcePrice(config, "text_total", keys...); value > 0 {
+			return value
+		}
+	}
 	if resource == "image_edit" {
 		return resourcePrice(config, "image", keys...)
 	}

apps/api/internal/runner/recording.go

@@ -151,6 +151,10 @@ func usageToMap(usage clients.Usage) map[string]any {
 		out["inputTokens"] = usage.InputTokens
 		out["promptTokens"] = usage.InputTokens
 	}
+	if usage.CachedInputTokens > 0 {
+		out["cachedInputTokens"] = usage.CachedInputTokens
+		out["cachedPromptTokens"] = usage.CachedInputTokens
+	}
 	if usage.OutputTokens > 0 {
 		out["outputTokens"] = usage.OutputTokens
 		out["completionTokens"] = usage.OutputTokens