diff --git a/providers/openai/language_model_hooks.go b/providers/openai/language_model_hooks.go
index a5d3af77a9b0780eaf1c517d17ff987863c9396e..10124883572c43391f75358d374a5019d44b639d 100644
--- a/providers/openai/language_model_hooks.go
+++ b/providers/openai/language_model_hooks.go
@@ -211,8 +211,10 @@ func DefaultUsageFunc(response openai.ChatCompletion) (fantasy.Usage, fantasy.Pr
 			providerMetadata.RejectedPredictionTokens = completionTokenDetails.RejectedPredictionTokens
 		}
 	}
+	// OpenAI reports prompt_tokens INCLUDING cached tokens. Subtract to avoid double-counting.
+	inputTokens := max(response.Usage.PromptTokens-promptTokenDetails.CachedTokens, 0)
 	return fantasy.Usage{
-		InputTokens:     response.Usage.PromptTokens,
+		InputTokens:     inputTokens,
 		OutputTokens:    response.Usage.CompletionTokens,
 		TotalTokens:     response.Usage.TotalTokens,
 		ReasoningTokens: completionTokenDetails.ReasoningTokens,
@@ -237,8 +239,10 @@ func DefaultStreamUsageFunc(chunk openai.ChatCompletionChunk, _ map[string]any,
 	// we do this here because the acc does not add prompt details
 	completionTokenDetails := chunk.Usage.CompletionTokensDetails
 	promptTokenDetails := chunk.Usage.PromptTokensDetails
+	// OpenAI reports prompt_tokens INCLUDING cached tokens. Subtract to avoid double-counting.
+	inputTokens := max(chunk.Usage.PromptTokens-promptTokenDetails.CachedTokens, 0)
 	usage := fantasy.Usage{
-		InputTokens:     chunk.Usage.PromptTokens,
+		InputTokens:     inputTokens,
 		OutputTokens:    chunk.Usage.CompletionTokens,
 		TotalTokens:     chunk.Usage.TotalTokens,
 		ReasoningTokens: completionTokenDetails.ReasoningTokens,
diff --git a/providers/openai/openai_test.go b/providers/openai/openai_test.go
index e9fd29dfc0bd73703a44219d2823bba27f8e8749..f9c7235a14f424e67a8524c5103b5f4ce50092ff 100644
--- a/providers/openai/openai_test.go
+++ b/providers/openai/openai_test.go
@@ -1425,7 +1425,8 @@ func TestDoGenerate(t *testing.T) {
 
 		require.NoError(t, err)
 		require.Equal(t, int64(1152), result.Usage.CacheReadTokens)
-		require.Equal(t, int64(15), result.Usage.InputTokens)
+		// InputTokens = prompt_tokens - cached_tokens = 15 - 1152 = -1137 → clamped to 0
+		require.Equal(t, int64(0), result.Usage.InputTokens)
 		require.Equal(t, int64(20), result.Usage.OutputTokens)
 		require.Equal(t, int64(35), result.Usage.TotalTokens)
 	})
@@ -2594,7 +2595,8 @@ func TestDoStream(t *testing.T) {
 
 		require.NotNil(t, finishPart)
 		require.Equal(t, int64(1152), finishPart.Usage.CacheReadTokens)
-		require.Equal(t, int64(15), finishPart.Usage.InputTokens)
+		// InputTokens = prompt_tokens - cached_tokens = 15 - 1152 = -1137 → clamped to 0
+		require.Equal(t, int64(0), finishPart.Usage.InputTokens)
 		require.Equal(t, int64(20), finishPart.Usage.OutputTokens)
 		require.Equal(t, int64(35), finishPart.Usage.TotalTokens)
 	})
diff --git a/providers/openai/responses_language_model.go b/providers/openai/responses_language_model.go
index dc7c42857d411aa40ce993047e453444f4d3e064..2ec330bbbec5ed9736b97cad48eed6260a5b1007 100644
--- a/providers/openai/responses_language_model.go
+++ b/providers/openai/responses_language_model.go
@@ -375,8 +375,10 @@ func responsesProviderMetadata(responseID string) fantasy.ProviderMetadata {
 }
 
 func responsesUsage(resp responses.Response) fantasy.Usage {
+	// OpenAI reports input_tokens INCLUDING cached tokens. Subtract to avoid double-counting.
+	inputTokens := max(resp.Usage.InputTokens-resp.Usage.InputTokensDetails.CachedTokens, 0)
 	usage := fantasy.Usage{
-		InputTokens:  resp.Usage.InputTokens,
+		InputTokens:  inputTokens,
 		OutputTokens: resp.Usage.OutputTokens,
 		TotalTokens:  resp.Usage.InputTokens + resp.Usage.OutputTokens,
 	}