diff --git a/public/src/chat.js b/public/src/chat.js
index dfd869e..e2e37f9 100644
--- a/public/src/chat.js
+++ b/public/src/chat.js
@@ -5,9 +5,9 @@ import { API_KEY, getAPIConfig } from "./config.js";
 import { getSelectedModel, getSelectedModelEndpoint, getSelectedModelMaxTokens, getSelectedModelMaxCompletionTokens } from "./models.js";
 import { groqAwareStreamingFetch } from "./groq-fetch.js";
 
-// Rough token estimation (4 chars = 1 token average for English)
+// Rough token estimation (3 chars = 1 token to be conservative)
 function estimateTokens(text) {
-	return Math.ceil(text.length / 4);
+	return Math.ceil(text.length / 3);
 }
 
 // Calculate available output tokens
diff --git a/public/src/translation.js b/public/src/translation.js
index 58efa0b..1410176 100644
--- a/public/src/translation.js
+++ b/public/src/translation.js
@@ -75,7 +75,7 @@ async function* sendMessageWithHistory(messageHistory) {
 	// Calculate available tokens for output based on current prompt
 	const modelMaxTokens = getSelectedModelMaxTokens();
 	const inputText = messageHistory.map(msg => msg.content).join('');
-	const inputTokens = Math.ceil(inputText.length / 4); // Simple token estimation
+	const inputTokens = Math.ceil(inputText.length / 3); // Conservative token estimation
 	const outputReserve = Math.floor(modelMaxTokens * 0.4); // Reserve 40% for output
 	let availableTokens = Math.max(512, modelMaxTokens - inputTokens - outputReserve);
 	
diff --git a/public/src/tts.js b/public/src/tts.js
index 0bdb657..1c42518 100644
--- a/public/src/tts.js
+++ b/public/src/tts.js
@@ -143,7 +143,7 @@ Your task:
 7. Start with the article title, then the main content
 
 Return only the article text ready for TTS, nothing else.`;
-	const inputTokens = Math.ceil((systemPrompt + preCleanedContent).length / 4);
+	const inputTokens = Math.ceil((systemPrompt + preCleanedContent).length / 3);
 	const buffer = Math.max(2048, Math.floor(inputTokens * 1.5)); // Use 1.5x input tokens as buffer, minimum 2048
 	const calculatedAvailable = Math.max(100, modelMaxTokens - inputTokens - buffer);
 	
diff --git a/public/src/ui.js b/public/src/ui.js
index b2ba880..c1a02ed 100644
--- a/public/src/ui.js
+++ b/public/src/ui.js
@@ -141,9 +141,9 @@ async function* sendMessageWithCustomHistory(messageHistory) {
 	const { getSelectedModelMaxTokens, getSelectedModelMaxCompletionTokens } = await import("./models.js");
 	const modelMaxTokens = getSelectedModelMaxTokens();
 	
-	// Estimate input tokens for intro generation
+	// Estimate input tokens for intro generation (3 chars = 1 token to be conservative)
 	const inputText = messageHistory.map(msg => msg.content).join('');
-	const inputTokens = Math.ceil(inputText.length / 4);
+	const inputTokens = Math.ceil(inputText.length / 3);
 	
 	// For intro generation, be more generous - we want complete intros
 	// With large context models, we have plenty of room