How NanoGPT Works, Step by Step

Most AI services explain their philosophy before they explain their mechanics. This walkthrough does the reverse. Here is what actually happens, in order, when you start using NanoGPT for the first time, from an empty browser session to a finished conversation with a receipt attached.

Step 1: Open the Site and Look Around

You land on a page that asks one question: "How can I help you?" Below the NanoGPT greeting sits a single input box with an "Ask anything" placeholder, a model selector reading "Chat · Free model", and a short line of copy: "ChatGPT, Claude, Gemini and hundreds more AI models in one place. Pay less, keep your chats on your device, and always access the newest models." Nothing on the NanoGPT home page asks for a credit card before you have seen the product.

The navigation is similarly spare: Models, Pricing, FAQ, Balance, Subscription. A NanoGPT visitor can browse the full model list and the full price table before creating any kind of account. That order matters, because the main NanoGPT proposition is the catalog itself, and the site lets you inspect the whole thing first.

Step 2: Create an Account and Add a Balance

Sign-up on NanoGPT is the standard email flow. The first real decision is the deposit. NanoGPT balances start at $0.10 if you pay in cryptocurrency or $1 if you pay by card, and the NanoGPT pricing page states there are no deposit fees. The minimum is deliberately small: the service wants you to test real usage with real money rather than commit to a plan.

Token billing is how NanoGPT charges that balance. The site gives a plain translation, roughly four tokens for every three words, and bills input and output separately. So a conversation that reads a long document costs more input than a one-line question, and a long NanoGPT answer costs more output than a short one. The meter starts when the first request does, and every NanoGPT thread shows its running total as it goes, so the balance is never a surprise at the end of a month.

Step 3: Pick a Model From the Catalog

This is the step that separates NanoGPT from a single-provider chat app. The NanoGPT models page organizes everything by modality: 652 text models, 231 image models, 187 video models, 109 audio models, 15 models for 3D, and a 20-model "decisions" category that the NanoGPT API exposes but the web interface does not. The counts come from 119+ AI labs, and 113 models were added in the month of this writing.

Choosing among them is where the NanoGPT benchmark pages earn their keep. NanoGPT tracks 114 benchmarks from 8 independent sources and publishes a composite score for 288 models. The catalog also surfaces task-based picks for coding, creative writing, roleplay, and math, ranked by the NanoGPT Composite score. A NanoGPT beginner can start from those lists instead of guessing whether a name they do not recognize is worth tokens.

Step 4: Chat, Code, or Create

With a model chosen, the NanoGPT input box behaves like any modern chat interface. You type, the model answers, and the conversation continues in threads. The NanoGPT home page summarizes the range plainly: chat, write, code, and create images and video. History for those threads stays on your device, which is a stated design decision rather than a settings toggle you have to find.

Switching models mid-session is where the NanoGPT workflow pays off. When a NanoGPT answer is close but not right, you can select a different family, from OpenAI or Anthropic or Google, from Qwen or DeepSeek or Mistral, and ask again against the same context. On NanoGPT, the switch is a dropdown change, not a new account on another site with another billing relationship to manage.

Step 5: Read What the Call Cost, Then Decide About Subscribing

Every NanoGPT response includes the exact cost charged for that request. This is the step most workflows skip, and it changes behavior in a small but real way. A NanoGPT session where each answer shows its price makes the tradeoff between a cheap model and a frontier model visible instead of theoretical.

The same transparency extends to the NanoGPT API. Every API response carries the exact cost of the call, so a script can sum its own spending. NanoGPT states that API prices are list prices with no markup, and gives a concrete example: a model priced at $5 per million input tokens and $30 per million output tokens at its original provider costs exactly $5 and $30 here.

Pay-as-you-go is the default NanoGPT mode, but heavy users can switch to a $12 per month subscription. The NanoGPT plan includes 60 million input tokens per week, free output on covered open-source models such as GLM, DeepSeek, Kimi, MiniMax, MiMo, and Qwen, and 100 images per day on four image models. The allowance resets every Monday at 00:00 UTC.

The subscription does not cover everything. Closed models such as Claude, GPT, Gemini, and Grok still bill from the balance, as do video, voice, transcription, web search, and some other tools. NanoGPT subscribers get a 5% discount on eligible paid text models, while pinning a specific provider adds a 5% surcharge, and bring-your-own-key requests are billed at 5% of the normal cost. The arithmetic only works if your usage matches the covered list.

Where the Steps Lead

Following the six steps in order, a NanoGPT account goes from empty to useful in about ten minutes: browse, sign up, deposit a dollar, pick from the task-based lists, ask, read the cost, and only then consider a subscription. The NanoGPT design puts the commitment at the end of the funnel rather than the beginning, which is the honest order for a prepaid service. Anyone who wants to verify the flow before trying it can read the NanoGPT pricing and models pages without an account, then decide whether the NanoGPT approach fits how they already work.