Why Sending Only “Hi” Can Still Consume Thousands of Tokens

[admin]

<h2><strong>Introduction</strong></h2><p></p><p>Recently, some users noticed an unusual situation:</p><blockquote><p>“I only sent ‘Hi’, why did it consume 10,000+ tokens?”</p><p></p></blockquote><p>At first glance, this may seem unexpected. A two-character message should only require a few tokens, right?</p><p>The answer is: <strong>the model does not only process your visible message.</strong></p><p>Modern AI assistants, especially coding agents such as Codex, Claude Code, and other IDE-based AI tools, send much more information together with your message to help the model understand the task and provide better responses.</p><p></p><hr><p></p><h2>Your Message Is Only a Small Part of the Context</h2><p>When you send:</p><pre class="rounded-lg bg-slate-950 px-4 py-3 text-slate-100"><code>Hi</code></pre><p></p><p>the actual request sent to the AI model may look more like this:</p><pre class="rounded-lg bg-slate-950 px-4 py-3 text-slate-100"><code>System Instructions
+
User Settings
+
Configured Skills
+
Available Tools
+
Project Information
+
File Context
+
Previous Conversation History
+
Your Message: "Hi"</code></pre><p></p><p>The total size of all these components is the actual context processed by the model.</p><p>Therefore, even a very short message can consume a significant number of tokens.</p><hr><p></p><h2><strong>What Can Increase Context Usage?</strong></h2><h3>1. Configured Skills</h3><p></p><p>If you have enabled custom skills or AI capabilities, their instructions need to be loaded into the model context.</p><p></p><p>For example:</p><pre class="rounded-lg bg-slate-950 px-4 py-3 text-slate-100"><code>Skill A instructions: 2,000 tokens
Skill B instructions: 3,000 tokens
Skill C instructions: 1,500 tokens</code></pre><p>Even before you type anything, these instructions may already occupy several thousand tokens.</p><hr><p></p><h3>2. Tools and Agent Capabilities</h3><p></p><p>AI coding assistants usually provide tools such as:</p><ul><li><p>File reading</p></li><li><p>Code search</p></li><li><p>Terminal execution</p></li><li><p>Git operations</p></li><li><p>Browser access</p></li><li><p>External integrations</p></li></ul><p>The definitions and usage instructions of these tools also consume context space.</p><hr><p></p><h3>3. Project and File Context</h3><p></p><p>When working inside an IDE or code repository, the AI client may automatically include:</p><ul><li><p>Project structure</p></li><li><p>Open files</p></li><li><p>Code snippets</p></li><li><p>Dependencies</p></li><li><p>Configuration files</p></li></ul><p>This allows the AI to understand your project, but it also increases context usage.</p><hr><p></p><h3>4. Previous Conversation History</h3><p></p><p>If you continue an existing conversation, the previous messages are usually included again so the model understands the context.</p><p>For example:</p><pre class="rounded-lg bg-slate-950 px-4 py-3 text-slate-100"><code>Previous discussion: 12,000 tokens

New message:
Hi</code></pre><p>The new request may still contain the previous 12,000 tokens.</p><hr><h2>How to Test Minimum Token Usage</h2><p>If you want to measure the token usage of a simple message, please try the following:</p><ol><li><p>Create a completely new conversation.</p></li><li><p>Disable custom skills.</p></li><li><p>Do not attach files.</p></li><li><p>Do not open a project workspace.</p></li><li><p>Send only:</p></li></ol><pre class="rounded-lg bg-slate-950 px-4 py-3 text-slate-100"><code>Hi</code></pre><p>This will show the approximate minimum context usage.</p><hr><h2>Context Window vs Token Usage</h2><p>It is important to distinguish between:</p><h3>Context Window</h3><p>The maximum amount of information a model can process in one request.</p><p>Example:</p><pre class="rounded-lg bg-slate-950 px-4 py-3 text-slate-100"><code>1M token context window</code></pre><p>means the model can theoretically handle up to 1 million tokens.</p><h3>Token Usage</h3><p>The actual amount of tokens included in a specific request.</p><p>Example:</p><pre class="rounded-lg bg-slate-950 px-4 py-3 text-slate-100"><code>User message:
Hi

Actual request:
System instructions + tools + history + "Hi"

Total:
15,000 tokens</code></pre><p>A large context window does not mean every request will only consume the length of your message.</p><p></p><hr><p></p><h2><strong>Conclusion</strong></h2><p></p><p>High token usage from a short message does not necessarily indicate a problem with the model or API.</p><p></p><p>In AI agent environments, the total context includes many hidden components required for intelligent operation. A simple “Hi” can still result in thousands of tokens being processed if the session contains skills, tools, project information, or previous history.</p><p></p><p>We recommend starting a clean session without additional context when testing raw token usage.</p><p></p><p>Thank you for your understanding and continued support.</p>