navveenb Token Counter Skill
When you find an LLM API call in code, estimate its token usage using these rules:
Estimation Rules
Input Tokens
- System prompt: Look for system message content. Estimate 1 token per 4 characters.
- User message: Look for user/human message content. If it includes variable interpolation, estimate based on typical content size.
- Few-shot examples: Count examples in the messages array. Each example ≈ 200-500 tokens.
- Tool descriptions: Count tool/function definitions. Each tool ≈ 100-500 tokens.
Output Tokens
- max_tokens parameter: If set, use that as the upper bound.
- If not set: Estimate by task type:
- Classification/routing: ~10-50 tokens
- Extraction: ~50-200 tokens
- Summarization: ~100-500 tokens
- Code generation: ~200-2000 tokens
- Creative writing: ~500-4000 tokens
Cost Rates (per 1M tokens)
| Model | Input | Output |
|---|---|---|
| Haiku / GPT-4o-mini | $0.25 | $1.25 |
| Sonnet / GPT-4o | $3.00 | $15.00 |
| Opus / GPT-4 / o1 | $15.00 | $75.00 |
Per Call Output
- Estimated input tokens: ____
- Estimated output tokens: ____
- Current model cost per call: $____
- Right-sized model cost per call: $____