How does NumStack calculate AI token costs?

NumStack uses real published pricing from OpenAI, Anthropic, Google, and other AI providers. Enter your typical usage (input tokens, output tokens, requests per day) to see projected monthly costs across all models.

Why do different AI models have different token prices?

Different models have different computational costs. Smaller models like GPT-4o mini or Haiku are cheaper but less capable. Larger models like GPT-4o or Claude Sonnet are more powerful but cost more per token.

How can I reduce my AI API costs?

Use smaller models for simple tasks, cache repeated prompts, trim output length (output tokens cost 3-5x more than input), and batch requests instead of real-time when latency allows.

← Back to NumStack

Prompt Optimizer

Reduce token usage without losing quality. Paste your prompt, choose your model, and get 3 optimized variants instantly.

Why prompt optimization saves real money

Input tokens are often overlooked in AI cost calculations — teams focus on output length but forget that long system prompts and verbose user instructions add up at scale. At 10,000 requests/day, even a 20% token reduction on a medium-sized prompt can save hundreds of dollars per month.

The Prompt Optimizer identifies the most common token-wasting patterns: filler phrases like "please can you", redundant qualifiers, duplicate sentences, and verbose vocabulary. It produces three variants so you can choose the right trade-off for your use case.

Three optimization strategies

🔵Remove Redundancy — Strips filler openers, redundant qualifiers, and duplicate sentences. Safest option that rarely changes meaning.
🟣Restructure for Clarity — Reorders the prompt to lead with the core task, removes elaboration based on your quality setting. Better for long, multi-paragraph prompts.
🔵Simplify Language — Replaces Latinate and multi-word phrases with direct alternatives. "Utilize" → "use", "in order to" → "to". Typically the highest reduction.

How accurate is the token counter?

The live token counter uses a 4-chars/token approximation, which matches GPT-family tokenization within ±5% for typical English prose. For exact counts, use the OpenAI tokenizer or the OpenAI Tokenizer playground. The approximation is accurate enough for cost estimates and optimization comparisons.

🔌 Want to automate this? Use the NumStack API to optimize prompts programmatically: POST /api/calculators/prompt-optimizer

How this calculator works

The optimizer estimates your prompt's token count using a 4-characters-per-token approximation, then runs three independent rewriting strategies against it and re-counts each result. The variant with the largest reduction is used to estimate monthly savings: it multiplies the tokens saved by your selected model's current input price and an assumed call volume, shown in the results, so you see an estimated dollar impact rather than just a token count.

Estimated monthly savings

tokensSaved = originalTokens − bestVariantTokens

monthlySavings = (tokensSaved ÷ 1,000,000) × inputPrice × callsPerMonth

How to use it

1.Paste the prompt you currently send to your AI model into the text area.
2.Select the model you use and how much quality you're willing to trade for savings.
3.Review the three optimized variants: redundancy removal, restructured for clarity, and simplified language.
4.Compare before/after token counts and estimated monthly savings, then copy the version you want to use.

Common mistakes

⚠Picking the most aggressive variant blindly — the biggest token reduction isn't always the safest choice — test the rewritten prompt against real inputs before shipping it, especially for tasks sensitive to exact instruction wording.
⚠Ignoring the quality setting — restructuring behaves differently depending on whether you marked the task accuracy-critical or speed-priority — mismatching it can strip constraints your prompt actually needs.
⚠Treating the token counter as exact — the 4-characters-per-token approximation is close but not exact — for billing-critical decisions, verify with your provider's real tokenizer.
⚠Optimizing once and never revisiting — small unnoticed additions creep back into prompts as they evolve — re-run the optimizer periodically instead of treating it as a one-time pass.

Frequently asked questions

Will optimizing my prompt change the AI's output?▾

It can. Removing redundancy is the safest option since it rarely changes meaning. Restructuring and simplifying language are more aggressive and can occasionally shift emphasis, so always test an optimized prompt against a few real examples before switching over in production.

How is the monthly savings estimate calculated?▾

It multiplies the token difference between your original prompt and the best-performing variant by your selected model's current input price and an assumed monthly call volume shown in the results. Your real savings depend on your actual call volume, which may differ from that default assumption.

Which of the three strategies should I use?▾

Remove Redundancy is the safest starting point since it rarely changes meaning. Restructure for Clarity works best on long, multi-paragraph prompts. Simplify Language typically produces the largest token reduction but changes wording the most — pick based on how much risk you're comfortable taking.

Does this work well for non-English prompts?▾

The token-counting approximation and vocabulary simplification rules are tuned for English text. Non-English prompts will still get a token estimate, but the simplify-language substitutions may not apply meaningfully.

Can I optimize system prompts, not just user messages?▾

Yes — paste any text you send to a model, including system prompts, few-shot examples, or full conversation templates. System prompts often have the most redundancy to trim since they're written once and rarely revisited.