See where your tokens go.
Live demo (coming soon)
Paste an LLM prompt. See a token-by-token cost heatmap. Find out which parts of your prompt are expensive -- and which models are cheapest.
LLM sticker shock is real. Token counters give you a number, but a number doesn't tell you where the cost is hiding. Is it the system prompt? The few-shot examples? That JSON blob you pasted in?
fntokenxray gives you a heatmap. Every word in your prompt is colored by its token density -- green for efficient, red for expensive. You see the shape of your costs at a glance.
- Token count -- estimated total for your prompt
- Cost estimate -- at the selected model's published input pricing
- Heatmap -- every word colored green-to-red by token density (tokens per visible character)
- Model comparison table -- same prompt, every model, sorted by cost
| Model | Provider | Input / 1M tokens |
|---|---|---|
| Claude Opus 4.6 | Anthropic | $15.00 |
| Claude Sonnet 4.6 | Anthropic | $3.00 |
| Claude Haiku 4.5 | Anthropic | $0.80 |
| GPT-4o | OpenAI | $2.50 |
| GPT-4o mini | OpenAI | $0.15 |
| Gemini 2.5 Pro | $1.25 |
Token counts are estimates, not exact. The tokenizer uses word-length heuristics (short words = 1 token, long words = 2-3 tokens, ~4 chars/token average). This is accurate enough for cost comparison and spotting expensive sections, but not a substitute for running text through an actual BPE tokenizer.
pnpm install
pnpm devBuild for production:
pnpm build
pnpm previewIf this tool saved you from a surprise API bill: