How to Estimate LLM API Costs Before You Build a Single Feature

Most teams discover their LLM bill is 3x what they expected — after the first month. The fix is a 5-minute estimate before you write a line of code. Here is the exact method that gets you within 20% of your real bill.

The Three Numbers You Need

Forget complicated spreadsheets. You only need three inputs to estimate any LLM workload:

1. Average input tokens per request — what you send to the model
2. Average output tokens per request — what the model sends back
3. Requests per day — your expected daily volume

Get those three numbers right, and you can predict any model's monthly bill in under a minute.

Step 1 — Estimate Input Tokens

Input tokens include everything you send: system prompt, conversation history, retrieved context (for RAG), and the user's current message. The big gotcha is conversation history — chatbots send the entire history with every message, which compounds fast.

Use case	Typical input tokens
Single-shot Q&A (no history)	100 - 500
Chatbot (5 turn history)	800 - 2,000
Chatbot (long history)	2,000 - 6,000
Document summarization (1 page)	1,500 - 3,000
Document summarization (10 pages)	15,000 - 30,000
RAG with 5 retrieved chunks	1,500 - 4,000
Code generation (with file context)	2,000 - 8,000
Long-context analysis (full doc)	20,000 - 100,000+

If you do not know yet, write a sample prompt and use the token counter to measure it. Multiply by your expected history length.

Step 2 — Estimate Output Tokens

Output is harder to estimate because it varies with the model, the prompt, and the task. Use these rough ranges:

Output type	Typical output tokens
Yes/no or single-word	5 - 20
Short answer (1-2 sentences)	30 - 80
Chat response (paragraph)	100 - 300
Detailed answer	300 - 800
Article or long-form	800 - 2,000
Code block (function)	200 - 800
Code block (full file)	800 - 3,000
Structured JSON (nested)	100 - 500

Tip: include "respond in under N words" or "limit response to N sentences" in your prompt to reduce output variance. This both lowers cost and tightens estimates.

Step 3 — Estimate Daily Volume

Be honest. Most projects estimate 10x more traffic than they actually get in month one. Use these starting points:

• Personal project / MVP: 10-100 requests/day
• Beta with 50 users: 200-1,000 requests/day
• Launched product (1k MAUs): 1,000-5,000 requests/day
• Growing product (10k MAUs): 10,000-50,000 requests/day
• Mid-market product (100k MAUs): 100,000+ requests/day

Multiply users by sessions per day per active user, then by requests per session. Most chat apps see 3-10 messages per session and 1-3 sessions per active user per day.

Step 4 — Pick Your Model and Run the Math

The formula is:

Monthly cost = ((input_tokens × input_price + output_tokens × output_price) ÷ 1,000,000) × requests_per_day × 30

Example: GPT-4o ($2.50 input, $10 output) at 1,500 input tokens, 400 output tokens, 2,000 requests per day:

• Input cost per request: 1,500 × $2.50 / 1,000,000 = $0.00375
• Output cost per request: 400 × $10 / 1,000,000 = $0.004
• Total per request: $0.00775
• Daily: $0.00775 × 2,000 = $15.50
• Monthly: $15.50 × 30 = $465

Same workload on GPT-4o mini: $0.000465 per request, $0.93 daily, $27.90 monthly. The cheap tier is roughly 17x less.

Step 5 — Add a 30% Buffer

Your real bill will exceed your estimate. Always. Reasons:

• Output tokens vary more than expected (some users ask for long responses)
• You will retry failed requests
• You will iterate on prompts during development (every test costs money)
• Conversation history grows over time as users chat more
• Edge cases hit higher token counts than averages suggest

Add 30% to your estimate. If the buffered number still fits your budget, you can build. If it doesn't, drop to a cheaper model or rework the prompt.

Skip the Math Entirely

The AI Cost Calculator does all of this in one input box. Type your three numbers, hit calculate, and see every major model side-by-side. The 6 built-in presets cover common workloads (chatbot, summarization, code gen, RAG, batch classification) so you can start from a realistic shape and adjust.