ChatGPT, Claude and every other AI don't count your words. They count tokens. Five short steps show where the money goes, then how to cut it. No technical background needed.
A small piece of text. Short words are one token, long words break into several. Type anything below and watch how the AI chops it up. Hindi usually takes more tokens than English for the same meaning.
Every message is billed twice. Input: what you send. Output: what the AI writes back. Output costs about 5 times more, because the AI reads everything at once but writes one token at a time. Drag the sliders and watch the receipt.
The AI has no memory. With every new message, the whole earlier chat is sent again, and billed again. By message 50 the AI is re-reading 49 old messages. Drag the slider.
Assumes 2,000 tokens of fixed instructions, 100-token questions, 300-token answers, on Claude Sonnet 5.5.
Take a real example: a support bot that sends a 50,000-token product manual with every question. Tap the cards on and off and watch the bill change.
Want to plug in your own numbers? Open the full calculator further down.
Per-token billing means more use, bigger bill, and one day the credits run dry. There are four ways around it: pool many free quotas, pay a flat monthly fee, use one key across many providers, or run the model on your own machine. Each has a catch.
An open-source gateway (MIT licence) that gives you one endpoint and pools many providers' free tiers behind it in a "combo". When one provider's quota runs out, the request moves to the next one automatically, with no error. The project's docs estimate up to 1.6 billion free tokens a month across 35 pools, though rate limits vary by provider.
npm install -g omniroute && omniroute # or docker run -p 20128:20128 diegosouzapw/omniroute
The dashboard opens on port 20128. Set INITIAL_PASSWORD first. In Claude Code, point ANTHROPIC_BASE_URL at the gateway. Source: github.com/diegosouzapw/OmniRoute
A subscription instead of per-token. You reserve one 8-hour daily window (Asia-Pacific, Europe or Americas) and get unlimited tokens on every model in your pool during it. Core is $22 a month (DeepSeek V4.1 Flash, MiMo V2.6 Flash), Frontier $71 (GLM 5.3, MiniMax M3), Flagship $199 (Kimi K3, Qwen3.8 Max). OpenAI and Anthropic SDK compatible, works with Claude Code.
Pick a plan at cheapestinference.com, choose your time window, create an API key, and change the base URL in your tool. Prices as of October 2026.
One API key, hundreds of models from many companies behind it. Pay as you go, but some models carry a ":free" tag and cost nothing within a rate limit. If one provider is down, it can fall back to another.
Create a key at openrouter.ai and filter the model list by "free". Check the current limits there.
Download open-source models and run them on your laptop or server. No credits, no per-token bill, and your data never leaves the machine. Pair it with Open WebUI for a ChatGPT-style interface.
ollama run llama3
Install from ollama.com. Providers like Google AI Studio and Groq also offer free tiers; check their current limits on their sites.
Three questions. Tap an answer. Prices are Claude Sonnet 5.5 at ₹96 per dollar.
Plug in your real numbers, test every lever on its own, and compare the same workload on every model.
| Model | Input | Output | Cache read | Per month |
|---|
Pick by job, not by brand: small models for FAQs, tagging and summaries; mid models for writing and coding; frontier models for hard reasoning.
MadeToScale × CodeBiosis set up caching, routing and batching for production AI apps.
contact@codebiosis.com