Apr 7, 2026
We get emails like this one fairly often:
Hi, I'm a novice regarding AI and prompting and just added $20 to my balance and started prompting. Can you give me some advice on what would be economical for me?
It's a totally fair question. AI pricing works differently from most software people are used to. There's no flat monthly fee (unless you want one), and costs scale with how much you use it. The good news is that this means you can keep costs very low — if you understand a few things about how it works.
Here's a practical guide to getting the most out of your NanoGPT balance.
On NanoGPT, you pay per token. A token is roughly ¾ of a word in English. Every time you send a message, you're billed for:
This means costs compound as conversations get longer. A chat that's 50 messages deep costs significantly more per message than a fresh one, because the model has to re-read all 50 previous messages every single time.
This is the single biggest reason people burn through balances unexpectedly. One long conversation on an expensive model can easily cost more than dozens of short ones on a cheap model.
Because the entire conversation history is sent with every message, the simplest thing you can do is start a new chat when you change topics.
If you spent 20 messages discussing a Python script and then want to switch to asking about a travel itinerary, open a new chat. There's no reason to keep paying to re-send the Python discussion to the model — it's not relevant anymore and it's costing you money.
Not all models cost the same. Not even close. Here's a rough sense of the range:
| Model tier | Examples | Relative cost |
|---|---|---|
| Very cheap | GPT 5.4 mini, Gemma 4, GLM 4.7 Flash | ~$0.15–0.30 per 1M input tokens |
| Mid-range | GPT 5.4, Gemini 3 Flash, Claude Sonnet 4.6 | ~$1–3 per 1M input tokens |
| Premium | Claude Opus 4.6, GPT 5.4 Pro, GLM 5.1 | ~$15–150+ per 1M input tokens |
The difference between the cheapest and most expensive models can be 1000x. For most everyday tasks — writing emails, brainstorming, asking questions, summarizing text — the cheap models are genuinely good enough. GPT 5.4 mini and Gemma 4 are excellent for the price.
Reserve the expensive models for when you actually need them: complex reasoning, nuanced code, legal or medical analysis, creative writing where quality really matters.
Practical rule of thumb: Start with a cheap model. If the answer isn't good enough, try a mid-range one. Only reach for the premium models when you have a specific reason to.
The NanoGPT subscription is $8/month and includes a range of models at no per-token cost. As of early 2026, subscription-included models include:
Open source has gotten really good. These models are competitive with the best closed-source models from a year ago, and they're 90–95% as capable as today's top models for most tasks. At $8/month with no per-token charges, the subscription is a genuine bargain if you use it regularly.
Even if you also use pay-per-token models for specialized tasks, the subscription can cover the bulk of your daily usage.
A 10-message chat on a mid-range model might cost you a few cents. A 100-message chat on the same model can cost dollars. The math is simple: if message 100 re-sends messages 1–99, you're paying for ~100x more input tokens than you were at message 1.
Tips:
Some models on NanoGPT support web search. Web search costs extra — roughly $0.01 per message on average on top of the regular token cost. This is worth it when you need up-to-date information, but don't leave it on for every message if you're just having a general conversation.
NanoGPT shows your remaining balance in the sidebar on the left. Keep an eye on it, especially when you're first getting a feel for costs. If you see it dropping fast, check which model you're using and how long your conversations are getting.
You can also set up auto-recharge so your balance automatically tops up when it gets low — but that's optional. Plenty of users just add credit manually when they need it.
Here's what typical usage actually costs on NanoGPT:
| Usage pattern | Model tier | Rough cost |
|---|---|---|
| 10 short messages | Cheap (e.g. GPT 5.4 mini) | < $0.01 |
| 10 short messages | Mid-range (e.g. GPT 5.4) | ~$0.05 |
| 10 short messages | Premium (e.g. Claude Opus 4.6) | ~$0.50 |
| 50 messages in a long chat | Mid-range | ~$0.50–1.00 |
| Full day of heavy use | Cheap | ~$0.10 |
| Full day of heavy use | Mid-range | ~$1–3 |
| Subscription (unlimited on included models) | GLM 5, Kimi K2.5, Deepseek V3.2, Minimax M2.7 | $8/month |
The range is enormous. The same person could spend $0.50/month or $50/month depending entirely on model choice and conversation management.
Most people who burn through their balance quickly are doing one of two things: using the most expensive models for everything, or letting conversations grow very long on paid models (or both). Fix those two things and costs drop dramatically.
If you're a casual user — a few dozen messages a day on mid-range models — expect to spend $2–5/month. If that sounds like you, the $8 subscription is probably the better deal, and you get more model for your money.
If you're a heavy user or developer running long sessions, be strategic about model choice and conversation hygiene. Start new chats, use cheap models for simple tasks, and reach for the heavy hitters only when you need them.
Questions? Email us at support@nano-gpt.com or open a ticket in our Discord. Happy to look at your usage and give personalized recommendations.
