Understanding Tokens
Your currency for AI. Learn how tokens work, what uses them, and how to maximize your budget.
What are tokens?
Tokens are units of text that Claude processes. Think of them as the building blocks of language. One token is roughly 4 characters in English text. When you ask Claude a question, both your input and Claude's response are measured in tokens.
Example: The phrase "Hello, how are you?" uses about 6 tokens.
You're charged based on:
- Input tokens: Every word in your message
- Output tokens: Every word Claude generates in the response
- Cache tokens: Tokens stored for faster, cheaper retrieval (10% the cost of input)
How tokens are used
Each interaction consumes tokens in a specific way:
| Action | Tokens Used | Notes |
|---|---|---|
| Asking a question | Input | Your message + conversation history |
| Claude's response | Output | Every word Claude generates |
| Uploaded file | Input | All text content is processed as input tokens |
| Code generation | Input + Output | Typically higher output token usage |
| Regenerate response | Input + Output | Full re-computation, same cost as new message |
What burns more tokens?
Some interactions consume significantly more tokens than others:
- Long conversations: Each new message includes previous conversation history, multiplying token usage
- Large files: Uploading PDFs, images, or code files counts every word
- Code-heavy requests: Asking Claude to write or debug large codebases
- Complex research: Multi-step reasoning and deep analysis use more output tokens
- Regenerations: Asking Claude to "try again" recalculates from scratch
- Web search: Each search adds tokens, plus the search results in your message
- Longer responses: Asking for detailed, verbose answers vs. concise ones
How to save tokens
Maximize your budget with these strategies:
- Start new chats strategically: Long conversation threads accumulate history. Start fresh when switching topics to avoid replaying context
- Be specific in requests: "Generate a function to sort an array" uses fewer tokens than rambling explanations
- Avoid regenerations: Refine your prompt instead of asking Claude to try again
- Use the tutor mode for learning: Haiku model (cheaper) is excellent for education, saving ~70% vs. Opus
- Paste only relevant code: Don't upload entire files; extract the relevant section
- Disable web search when unnecessary: Only enable it for time-sensitive queries
- Request concise responses: Ask for bullet points instead of essays when appropriate
- Consolidate multi-step tasks: Ask multiple related questions in one message vs. separate messages
Pricing by model
Different Claude models have different costs. Your plan determines your base model, but you can downgrade to cheaper models per-conversation:
- Haiku (Eco): Fastest, cheapest. Great for summarization, tutoring, simple questions
- Sonnet (Balanced): Default. Best balance of speed, cost, and capability
- Opus (Performance): Most capable but 5-10× more expensive. Use only for complex reasoning
Reading your token usage
In Aurelius, you'll see three numbers in your dashboard:
- Input tokens: Words you and Claude sent (including history)
- Cache tokens: Tokens retrieved from cache (10% cost)
- Output tokens: Words Claude generated
Your cost = (input × rate) + (cache × 0.1 × rate) + (output × rate). Rates vary by model.
Pro tips
- Use the "Code" mode in Aurelius to get code optimized for efficiency
- Request Claude to "save tokens by being concise" in your system preferences
- Delete old conversations to keep your sidebar lean (no impact on costs)
- Pair simple questions with Haiku, complex problems with Sonnet
- Ask Claude directly: "Summarize this in 50 tokens or less"