diff --git a/src/core/task/Task.ts b/src/core/task/Task.ts
index 08880d0ee0..bb0f549c5a 100644
--- a/src/core/task/Task.ts
+++ b/src/core/task/Task.ts
@@ -1920,9 +1920,13 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 									cacheReadTokens,
 								)
 
+					// For tokensIn, store the base input tokens (not including cache tokens)
+					// This ensures context length calculation works correctly in getApiMetrics
+					// where it adds tokensIn + tokensOut. The cache tokens are stored separately
+					// in cacheWrites and cacheReads for display purposes.
 					this.clineMessages[lastApiReqIndex].text = JSON.stringify({
 						...existingData,
-						tokensIn: costResult.totalInputTokens,
+						tokensIn: inputTokens,
 						tokensOut: costResult.totalOutputTokens,
 						cacheWrites: cacheWriteTokens,
 						cacheReads: cacheReadTokens,
diff --git a/src/shared/getApiMetrics.ts b/src/shared/getApiMetrics.ts
index 2ae4756764..4f13fefc65 100644
--- a/src/shared/getApiMetrics.ts
+++ b/src/shared/getApiMetrics.ts
@@ -82,9 +82,10 @@ export function getApiMetrics(messages: ClineMessage[]) {
 				const parsedText: ParsedApiReqStartedTextType = JSON.parse(message.text)
 				const { tokensIn, tokensOut } = parsedText
 
-				// Since tokensIn now stores TOTAL input tokens (including cache tokens),
-				// we no longer need to add cacheWrites and cacheReads separately.
-				// This applies to both Anthropic and OpenAI protocols.
+				// tokensIn contains base input tokens (without cache tokens).
+				// This works correctly for both Anthropic and OpenAI protocols because:
+				// - Cache tokens are stored separately in cacheWrites/cacheReads
+				// - Context length = base input + output tokens
 				result.contextTokens = (tokensIn || 0) + (tokensOut || 0)
 			} catch (error) {
 				console.error("Error parsing JSON:", error)