diff --git a/bridge_integration_test.go b/bridge_integration_test.go
index c2f307e..a5df3e0 100644
--- a/bridge_integration_test.go
+++ b/bridge_integration_test.go
@@ -736,11 +736,13 @@ func TestOpenAIInjectedTools(t *testing.T) {
 			require.Contains(t, content.Message.Content, "dd711d5c-83c6-4c08-a0af-b73055906e8c") // The ID of the workspace to be returned.
 
 			// Check the token usage from the client's perspective.
-			assert.EqualValues(t, 9911, message.Usage.PromptTokens)
+			// This *should* work but the openai SDK doesn't accumulate the prompt token details :(.
+			// See https://github.com/openai/openai-go/blob/v2.7.0/streamaccumulator.go#L145-L147.
+			// assert.EqualValues(t, 5047, message.Usage.PromptTokens-message.Usage.PromptTokensDetails.CachedTokens)
 			assert.EqualValues(t, 105, message.Usage.CompletionTokens)
 
 			// Ensure tokens used during injected tool invocation are accounted for.
-			require.EqualValues(t, 9911, calculateTotalInputTokens(recorderClient.tokenUsages))
+			require.EqualValues(t, 5047, calculateTotalInputTokens(recorderClient.tokenUsages))
 			require.EqualValues(t, 105, calculateTotalOutputTokens(recorderClient.tokenUsages))
 		})
 	}
diff --git a/intercept_openai_chat_blocking.go b/intercept_openai_chat_blocking.go
index 7cefc5c..b6a0343 100644
--- a/intercept_openai_chat_blocking.go
+++ b/intercept_openai_chat_blocking.go
@@ -81,7 +81,7 @@ func (i *OpenAIBlockingChatInterception) ProcessRequest(w http.ResponseWriter, r
 		_ = i.recorder.RecordTokenUsage(ctx, &TokenUsageRecord{
 			InterceptionID: i.ID().String(),
 			MsgID:          completion.ID,
-			Input:          lastUsage.PromptTokens,
+			Input:          calculateActualInputTokenUsage(lastUsage),
 			Output:         lastUsage.CompletionTokens,
 			Metadata: Metadata{
 				"prompt_audio":                   lastUsage.PromptTokensDetails.AudioTokens,
diff --git a/intercept_openai_chat_streaming.go b/intercept_openai_chat_streaming.go
index 314d20c..f10c26e 100644
--- a/intercept_openai_chat_streaming.go
+++ b/intercept_openai_chat_streaming.go
@@ -165,7 +165,7 @@ func (i *OpenAIStreamingChatInterception) ProcessRequest(w http.ResponseWriter,
 			_ = i.recorder.RecordTokenUsage(streamCtx, &TokenUsageRecord{
 				InterceptionID: i.ID().String(),
 				MsgID:          processor.getMsgID(),
-				Input:          lastUsage.PromptTokens,
+				Input:          calculateActualInputTokenUsage(lastUsage),
 				Output:         lastUsage.CompletionTokens,
 				Metadata: Metadata{
 					"prompt_audio":                   lastUsage.PromptTokensDetails.AudioTokens,
diff --git a/openai.go b/openai.go
index b296123..dc3abc8 100644
--- a/openai.go
+++ b/openai.go
@@ -96,6 +96,15 @@ func sumUsage(ref, in openai.CompletionUsage) openai.CompletionUsage {
 	}
 }
 
+// calculateActualInputTokenUsage accounts for cached tokens which are included in [openai.CompletionUsage].PromptTokens.
+func calculateActualInputTokenUsage(in openai.CompletionUsage) int64 {
+	// Input *includes* the cached tokens, so we subtract them here to reflect actual input token usage.
+	// The original value can be reconstructed by referencing the "prompt_cached" field in metadata.
+	// See https://platform.openai.com/docs/api-reference/usage/completions_object#usage/completions_object-input_tokens.
+	return in.PromptTokens /* The aggregated number of text input tokens used, including cached tokens. */ -
+		in.PromptTokensDetails.CachedTokens /* The aggregated number of text input tokens that has been cached from previous requests. */
+}
+
 func getOpenAIErrorResponse(err error) *OpenAIErrorResponse {
 	var apierr *openai.Error
 	if !errors.As(err, &apierr) {