diff --git a/billing.go b/billing.go
index b53ddea..194617e 100644
--- a/billing.go
+++ b/billing.go
@@ -56,14 +56,25 @@ func CalculateCost(provider, model string, costInfo CostInfo, promptTokens, comp
 		cachedTokens = cacheUsage.CachedTokens + cacheUsage.CacheReadInputTokens
 	}
 
-	// Clamp to prompt tokens as a safety net — cached tokens can never
-	// exceed the total prompt tokens the provider reported.
+	// Providers report prompt tokens differently:
+	//   - OpenAI/Fireworks/Bedrock: promptTokens INCLUDES cached tokens
+	//     → non-cached = promptTokens - cachedTokens
+	//   - Anthropic: input_tokens EXCLUDES cached tokens (only new tokens)
+	//     → non-cached = promptTokens (as reported), cached is additional
+	//
+	// We detect the style by comparing: if cached > prompt, the provider
+	// must be reporting non-cached only (Anthropic style).
+	var nonCachedTokens int
 	if cachedTokens > promptTokens {
-		cachedTokens = promptTokens
+		// Anthropic style: promptTokens = non-cached only, cached is separate
+		nonCachedTokens = promptTokens
+		// Adjust promptTokens to reflect the true total for the BillingResult
+		promptTokens = promptTokens + cachedTokens
+	} else {
+		// OpenAI style: promptTokens includes cached
+		nonCachedTokens = promptTokens - cachedTokens
 	}
 
-	nonCachedTokens := promptTokens - cachedTokens
-
 	// Non-cached prompt tokens at full input rate
 	inputCost := costInfo.Input * float64(nonCachedTokens) / 1_000_000
 
diff --git a/billing_test.go b/billing_test.go
index de448d2..72ce02a 100644
--- a/billing_test.go
+++ b/billing_test.go
@@ -67,19 +67,27 @@ func TestCalculateCost_WithOpenAICacheHit(t *testing.T) {
 }
 
 func TestCalculateCost_WithAnthropicCacheHit(t *testing.T) {
+	// Anthropic reports input_tokens as non-cached only (excludes cached).
+	// Real example: input_tokens=3, cache_read_input_tokens=2022
 	costInfo := CostInfo{Input: 3.0, Output: 15.0, CacheRead: 0.3}
-	cacheUsage := &CacheUsage{CacheReadInputTokens: 2000}
+	cacheUsage := &CacheUsage{CacheReadInputTokens: 2022}
 
-	result := CalculateCost("anthropic", "claude-sonnet-4", costInfo, 2500, 100, cacheUsage)
+	// promptTokens=3 (Anthropic's input_tokens, non-cached only)
+	result := CalculateCost("anthropic", "claude-sonnet-4", costInfo, 3, 10, cacheUsage)
 
-	if result.CachedTokens != 2000 {
-		t.Errorf("CachedTokens = %d, want 2000", result.CachedTokens)
+	if result.CachedTokens != 2022 {
+		t.Errorf("CachedTokens = %d, want 2022", result.CachedTokens)
 	}
 
-	// 500 non-cached at full rate, 2000 cached at cache rate
-	expectedInput := 3.0 * 500 / 1_000_000
-	expectedCached := 0.3 * 2000 / 1_000_000
-	expectedOutput := 15.0 * 100 / 1_000_000
+	// PromptTokens should be adjusted to include cached: 3 + 2022 = 2025
+	if result.PromptTokens != 2025 {
+		t.Errorf("PromptTokens = %d, want 2025 (non-cached + cached)", result.PromptTokens)
+	}
+
+	// 3 non-cached at full rate, 2022 cached at cache rate
+	expectedInput := 3.0 * 3 / 1_000_000
+	expectedCached := 0.3 * 2022 / 1_000_000
+	expectedOutput := 15.0 * 10 / 1_000_000
 
 	assertFloat(t, "InputCost", result.InputCost, expectedInput)
 	assertFloat(t, "CachedInputCost", result.CachedInputCost, expectedCached)
@@ -105,21 +113,26 @@ func TestCalculateCost_CacheUsageWithZeroTokens(t *testing.T) {
 	assertFloat(t, "TotalCost", result.TotalCost, expectedInput+expectedOutput)
 }
 
-func TestCalculateCost_CacheUsageExceedsPromptTokens(t *testing.T) {
+func TestCalculateCost_CacheExceedsPromptTokens_AnthropicStyle(t *testing.T) {
+	// Anthropic style: prompt_tokens=10 (non-cached), cache_read=5000 (cached).
+	// cached > prompt means the provider reports non-cached only.
 	costInfo := CostInfo{Input: 3.0, Output: 15.0, CacheRead: 1.5}
-	// More cached tokens than prompt tokens (shouldn't happen, but defensive)
-	cacheUsage := &CacheUsage{CachedTokens: 5000}
+	cacheUsage := &CacheUsage{CacheReadInputTokens: 5000}
 
-	result := CalculateCost("openai", "gpt-4o", costInfo, 1000, 500, cacheUsage)
+	result := CalculateCost("anthropic", "claude-sonnet-4", costInfo, 10, 500, cacheUsage)
 
-	// Cached tokens should be clamped to prompt tokens
-	if result.CachedTokens != 1000 {
-		t.Errorf("CachedTokens = %d, want 1000 (clamped)", result.CachedTokens)
+	if result.CachedTokens != 5000 {
+		t.Errorf("CachedTokens = %d, want 5000", result.CachedTokens)
+	}
+	// PromptTokens adjusted to true total: 10 + 5000
+	if result.PromptTokens != 5010 {
+		t.Errorf("PromptTokens = %d, want 5010", result.PromptTokens)
 	}
 
-	// All prompt tokens at cache rate, none at full rate
-	assertFloat(t, "InputCost", result.InputCost, 0)
-	expectedCached := 1.5 * 1000 / 1_000_000
+	// 10 non-cached at full rate, 5000 cached at cache rate
+	expectedInput := 3.0 * 10 / 1_000_000
+	expectedCached := 1.5 * 5000 / 1_000_000
+	assertFloat(t, "InputCost", result.InputCost, expectedInput)
 	assertFloat(t, "CachedInputCost", result.CachedInputCost, expectedCached)
 }