Token cost per output unit calculator
Per-token pricing is meaningless to most of the people who approve AI budgets. This calculator translates it into the units your business actually ships — a blog post, a support reply, a code review — so you can say “each one costs half a cent” instead of quoting prices per million tokens.
The communication problem with token pricing
When an engineer says “GPT-4o costs $5 per million input tokens and $15 per million output tokens,” a product manager or CFO has no intuitive frame for whether that is cheap or expensive. Is it affordable to generate 10,000 product descriptions? 500 personalized emails? A daily batch of support summaries? The token price does not tell you.
Unit cost answers the practical question: how much does it cost to produce one of the things this feature actually creates? That number travels well across teams and budget discussions.
How it works
The calculator converts your average output length from words to tokens, adds the prompt tokens you send to produce it, and applies the model’s separate input and output prices:
output_tokens = output_words × 1.3
unit_cost = prompt_tokens / 1M × input_price
+ output_tokens / 1M × output_priceIt then multiplies by 100 and 1,000 so you have batch figures ready for a spreadsheet or a budget request.
Common unit costs by output type
These are illustrative ranges based on typical prompt and output lengths for each format — not guaranteed figures. Your actual costs depend on your specific prompts, models, and output lengths.
| Output type | Typical prompt tokens | Typical output tokens | Notes |
|---|---|---|---|
| Short support reply | 500–800 | 100–200 | Context, history, and instructions in the prompt |
| Long-form blog post (~1,000 words) | 300–600 | 1,300 | One English word ≈ 1.3 tokens |
| Product description | 200–400 | 150–300 | Short, focused output |
| Code review comment | 1,500–4,000 | 200–500 | Large code file as input, concise output |
| Email draft | 400–700 | 300–600 | Variable depending on length |
Tips for accurate unit costing
- Measure real outputs. Sample 50 actual generations from your pipeline, count their words, and use that average — aspirational length assumptions understate cost.
- Measure real prompts. The system prompt plus any context you inject can easily exceed your visible “question” in token count. Log token counts from your API responses.
- Report the unit cost with its assumptions. Note the model, input length, and output length you assumed so the number is reproducible when models or prices change.
- Compare models per unit, not per token. A cheaper model may write 40% more words to say the same thing; cost per unit catches that overhead, cost per token does not.