NanoGPT for Roleplay (2026): Pricing, the $12 Subscription, and Setup for Janitor AI and SillyTavern
NanoGPT is a paid service that gives you one API key for hundreds of AI models, and you pay either per message from a prepaid balance or through a $12 monthly subscription. Roleplayers use it to bring a better model to Janitor AI, SillyTavern, Chub or RisuAI.
It is one of the three provider names that come up most when players ask where to get an API key. The other two are OpenRouter and Chutes. This guide covers what NanoGPT costs for roleplay, what the subscription really includes, how to connect it, and what goes wrong.
Everything here comes from NanoGPT’s own pricing page, documentation, privacy policy and public model lists, read on 11 October 2026. We did not open a paid account or send requests, so this page makes no claim about speed or writing quality.
One note on the name. nanoGPT is also a well-known code project by Andrej Karpathy for training small language models. That project has nothing to do with this service. If your search results are full of training code, you found the other one.
New to API keys? Start with how a Janitor AI API key works. Comparing providers? See OpenRouter vs NanoGPT vs Chutes.
The Short Answer
- It is pay-as-you-go by default. You add a balance and each message takes a small amount from it.
- The smallest deposit is $0.10 with crypto or $1 with a card, and NanoGPT says it charges no fee on deposits.
- The $12 subscription is optional. It includes 60 million input tokens a week on a list of open-weight models.
- Claude, GPT and Gemini are available, but only on the balance. They are not in the subscription.
- For a casual player, pay-as-you-go costs less than the subscription. We show the arithmetic below.
- It is not free for API use. The one free model works on the website only.
What NanoGPT Is
NanoGPT is an aggregator. It does not make AI models. It connects to the companies that run them and gives you one account, one balance and one key for all of them.
On 11 October 2026 its main model list had 643 entries. That includes the large commercial ones, such as Claude, GPT and Gemini, the large open-weight ones, such as DeepSeek, GLM, Kimi and Qwen, and many community fine-tunes that are made for roleplay and fiction.
Two things set it apart from other aggregators, according to its own pages:
- No account is required. The privacy policy says you can “use our services without creating an account”. Its FAQ says it “does not require KYC or an ID check”.
- Payment is flexible. It accepts cards and several cryptocurrencies, including one called Nano.
What NanoGPT Costs
Pay as you go
The pricing page says: “API at list prices — no markup”. It gives an example: a model that costs $5 and $30 per million tokens at its maker costs “exactly $5/$30 here”.
It names two exceptions, and both are things you choose to do:
- Pinning your requests to one specific provider “adds 5%”.
- Using your own provider key through NanoGPT is “billed 5% of the normal model cost”.
It also says there are “no percentage fees on deposits”, and that every reply from the API includes the exact amount charged, so you can check.
What a roleplay session costs
To make prices easy to compare, we use one example session on every provider page: 30 minutes of roleplay, about 25 turns, with 80,000 input tokens and 12,000 output tokens in total. The example is from Chutes’s pricing calculator. The arithmetic below is ours, using the prices in NanoGPT’s API on 11 October 2026.
| Model | Input, per 1M tokens | Output, per 1M tokens | One session | In the subscription |
|---|---|---|---|---|
| DeepSeek V4 Flash (0731) | $0.05 | $0.16 | $0.006 | Yes |
| DeepSeek V4.1 Flash | $0.13 | $0.52 | $0.017 | Yes |
| Qwen 3.8 27B | $0.15 | $0.70 | $0.020 | Yes |
| DeepSeek V3.2 | $0.28 | $0.42 | $0.027 | Yes |
| Hermes 4 405B | $0.30 | $1.20 | $0.038 | Yes |
| GLM 5.2 | $0.42 | $1.32 | $0.049 | Yes |
| GLM 5.3 | $0.60 | $1.10 | $0.061 | Yes |
| Kimi K2.6 | $0.50 | $2.60 | $0.071 | Yes |
| DeepSeek V4 Pro (0813) | $1.10 | $2.50 | $0.118 | Yes |
| Kimi K3 | $2.00 | $10.00 | $0.28 | No |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $0.28 | No |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.56 | No |
Two things to read from this.
The cheap models are very cheap. Half an hour on DeepSeek V4 Flash is under one cent. A player who does one session a day spends about 18 cents a month.
The premium models cost about fifty times more. The same half hour on Claude Sonnet 5.5 is 28 cents. That is still small, but a heavy month on Claude adds up in a way a month on DeepSeek does not. Our best Claude model for roleplay guide covers when the difference is worth paying for.
In both cases, most of the bill is input. Every turn sends the character card, the lorebook entries and the chat so far. Our token budget guide explains how to keep that number down.
The $12 Subscription
The pricing page calls the subscription Pro. On 11 October 2026 it listed:
- $12 a month.
- 60 million included input tokens a week.
- Text and image models included, with 100 images a day.
- A 5% discount on eligible paid text models.
- Use through both the website and the API.
What “60 million input tokens a week” means
The quota counts input tokens. NanoGPT’s documentation is clear on this: “Input-token quotas count tokens, not requests.” So the size of your context decides how far the quota goes, not the number of messages.
| Input sent per turn | Turns in a week | Turns in a day |
|---|---|---|
| 3,200 tokens (a new, short chat) | 18,750 | about 2,680 |
| 20,000 tokens | 3,000 | about 430 |
| 50,000 tokens | 1,200 | about 170 |
| 100,000 tokens (a very long chat with a large context) | 600 | about 85 |
For most players this is more than they will use. It gets tight only when you set a very large context and play a lot. A swipe or a regenerate sends the whole context again, so it counts in full each time.
NanoGPT’s usage API also reports a daily counter next to the weekly one. The pricing page gives only the weekly number. If you subscribe, the usage page in your account shows your real limits and when they reset. Trust those numbers over any guide, including this one.
What is in the included list, and what is not
NanoGPT publishes the list of subscription models. On 11 October 2026 it had 311 entries. We read it for the models roleplayers ask about.
Included:
- DeepSeek: V4 Flash, V4.1 Flash, V4 Pro and V3.2, with their thinking versions.
- GLM 5, 5.1, 5.2 and 5.3.
- Kimi K2.5 and K2.6.
- Qwen 3.5, 3.6 and 3.8.
- MiMo 2.5 and 2.6.
- Hermes 4 405B and NVIDIA’s Nemotron 3 Ultra.
- A large set of community roleplay fine-tunes, including several from TheDrummer (Cydonia, Anubis, Rocinante, UnslopNemo) and Sao10K (Euryale, Stheno, Hanami), and Magnum.
Not included:
- Claude, GPT and Gemini. We found no model from Anthropic, OpenAI or Google’s Gemini family in the list.
- Kimi K3.
You can still use those models on a subscribed account. They are charged to your balance at the normal price.
The list changes. NanoGPT’s documentation says to copy the “exact current model ID” from its pricing page, and the pricing page has a switch that shows only subscription models.
Is the subscription worth it?
Here is the comparison with pay-as-you-go, using the session costs from the table above.
| Model | Sessions that $12 buys on pay-as-you-go | Per day, over a month |
|---|---|---|
| DeepSeek V4 Flash (0731) | about 2,000 | about 67 |
| DeepSeek V3.2 | about 440 | about 15 |
| GLM 5.2 | about 245 | about 8 |
| Kimi K2.6 | about 170 | about 6 |
| DeepSeek V4 Pro (0813) | about 100 | about 3 |
Read it this way: if you play fewer sessions than the last column, pay-as-you-go is cheaper. If you play more, the subscription is cheaper.
- On the cheapest models, almost nobody reaches the line. 67 half-hour sessions a day is more than a day holds.
- On mid-priced models the line is 6 to 8 sessions a day. That is three to four hours of play, every day.
- On DeepSeek V4 Pro the line is about 3 sessions a day. A daily player crosses it.
So the honest answer is: start on pay-as-you-go. Put in a few dollars and look at what a week costs you. Subscribe only if your real spending is heading past $12 a month on models that are in the included list.
One more reason people subscribe is not about money. A subscription gives a fixed bill. If you would rather pay $12 and stop thinking about it, that is a fair choice. NanoGPT’s refund policy says you can cancel at any time to stop future renewals, and that cancelling does not refund the current period.
How to Connect NanoGPT
The two addresses
NanoGPT has two addresses, and guides mix them up.
| Address | What NanoGPT says it is for |
|---|---|
https://api.nano-gpt.com/api/v1 | The direct API. Recommended for API keys. It allows larger requests and longer running times |
https://nano-gpt.com/api/v1 | The website’s API. It works, with lower limits |
NanoGPT’s documentation says to use the direct address for “large attachments, long conversations, and model calls that take time to finish”. A long roleplay with a big context is a long conversation, so the direct address is the better default.
One warning from the same page: the two addresses “can have different model catalogs during deployment rollouts”. A model that is listed on one is not always available on the other. If a model returns “not found” on one address, check it on the other.
Janitor AI
These steps are from NanoGPT’s own Janitor AI guide.
- Get your API key at
nano-gpt.com/api. It starts withsk-nano-. - In Janitor AI, open API Settings.
- Select Proxy as the provider.
- Click + Add under Proxy Configurations.
- Fill in the four fields:
- Configuration Name: any label, for example NanoGPT.
- Model Name: the model ID, for example
deepseek/deepseek-v4-flash-0731. - Proxy URL:
https://api.nano-gpt.com/api/v1/chat/completions - API Key: your NanoGPT key.
- Click Apply, and make sure the NanoGPT configuration is marked Active.
If you subscribe, NanoGPT offers a shortcut for the model name: auto-model-basic. Its guide says this “automatically selects a model included in your subscription”. Use a specific model ID if you want to know which model is writing.
To see the model’s reasoning in Janitor AI, NanoGPT says to make a second configuration with a different address:
https://api.nano-gpt.com/api/v1thinking/chat/completions
With that address, the reasoning and the answer arrive together as normal text. With the standard address, they arrive separately, and Janitor AI shows only the answer. Most players want the standard one.
Remember that the character must allow proxy use. That is the creator’s setting, not yours. Our Janitor AI proxy guide covers that screen, and best model for Janitor AI helps you choose a model.
SillyTavern
SillyTavern 1.19.0 has NanoGPT built in.
- Open API Connections.
- Set the API to Chat Completion.
- Choose NanoGPT as the source.
- Paste your key and connect.
- Pick a model from the list.
SillyTavern’s code uses the website address (https://nano-gpt.com/api/v1) for this built-in source. It also has two NanoGPT options of its own: one to choose a provider, and one to force pay-as-you-go billing. Both have a cost, explained in the next section. Leave them alone unless you know why you want them.
NanoGPT is also one of SillyTavern’s image sources. Our SillyTavern image generation guide covers that. NanoGPT’s documentation adds one tip: SillyTavern starts at 512x512, and some image models reject that size, so choose a size the model supports.
Other apps
NanoGPT’s documentation has setup pages for RisuAI and several other apps. Any app with an “OpenAI-compatible” option can use the direct address and your key. For Chub, see how to use Chub AI.
Controlling What You Spend
NanoGPT has three controls that are useful for a roleplayer, and they are easy to miss.
A daily limit on each key
When you create or edit an API key, you can set a limit of requests per day, a limit of dollars per day, or both. The counters reset at midnight UTC. NanoGPT notes that the dollar limit uses an estimate, so the real cap is approximate.
This is the simplest protection there is. Make one key for Janitor AI, give it a limit of $1 a day, and a mistake cannot cost more than that. When the limit is reached, the API returns error 429 with a message such as “Daily request limit exceeded”.
The billing mode of each key
If you subscribe, each key has a billing setting with three choices: subscription first, subscription only, or pay-as-you-go only. The documentation says: “A key saved as Subscription only cannot spend your balance.”
For a subscriber who wants a fixed bill, subscription only is the setting that guarantees it. With that setting, choosing a model outside the included list fails. It does not charge you.
Choosing a provider costs extra
Some models are run by more than one company. NanoGPT picks one for you. You can choose one yourself, but the documentation says an explicit choice “is always pay-as-you-go” and is charged “including provider-selection markup”. For subscribers it goes further: choosing a provider “bypasses subscription coverage for that request”.
So if you are subscribed and you set a provider in SillyTavern, those messages come out of your balance, not your subscription. If your balance drops while you thought you were inside the subscription, check that setting first.
NanoGPT Errors and What They Mean
The status codes below are from NanoGPT’s error documentation.
| Code | Meaning | What to do |
|---|---|---|
| 401 | The API key is missing, wrong or revoked | Paste the key again. Check for spaces |
| 402 | Insufficient balance | Add balance, or choose a model that your subscription includes |
| 403 | The key is not allowed to do this | Check the key’s billing mode and the model |
| 404 | The model was not found | Copy the current model ID from the pricing page |
| 413 | The request is too large | Use the direct address, or lower the context |
| 429 | A rate limit or a daily key limit | Wait. If it is your own daily limit, it resets at midnight UTC |
| 503 | The model is temporarily unavailable | Try again, or change model |
”PROXY ERROR” in Janitor AI
Janitor AI often shows one general message, “no response from bot”, for any failure. NanoGPT’s Janitor guide lists the three usual causes:
- A paid model with no balance. You chose a model that is not in your subscription, and your balance is empty.
- A wrong model ID. The ID is “outdated, misspelled, or unavailable”. A 404 “usually means you should copy the model’s current ID”.
- A wrong address or key. Check the proxy URL letter by letter, and check that the key “has no extra spaces”.
The first one catches subscribers who pick Claude or Kimi K3 and expect the subscription to cover it.
A reply that stops in the middle
NanoGPT’s documentation says that an error can arrive after the reply has started. It also warns: “A stream failure does not by itself establish whether a charge was recorded.” In simple terms, a reply that fails halfway may or may not have been charged. Each reply has a request ID that support can look up.
Other causes
If the error is not in the table, it may come from the chat app. See Janitor AI not working and SillyTavern not working. NanoGPT’s general limit is 25 requests a second, which no single roleplayer will reach.
Privacy
From NanoGPT’s privacy policy, in its own words:
“By default, we do NOT store prompt or conversation content on our servers.”
On the website, the policy says conversation history is kept in “your local browser storage”. Storage on NanoGPT’s servers happens only if you turn on optional features such as conversation sync or sharing.
The same policy is direct about the other half:
“When you submit a prompt, it is passed directly to the relevant service provider… these service providers may store and process this information according to their own privacy policies.”
That is true of every aggregator, and NanoGPT says it plainly. The policy then lists what individual providers do, one by one. Some state zero retention. Some do not. For a model run by several companies, you do not always know which one served your message unless you choose it, and choosing costs extra.
NanoGPT also lists versions of some models with names that start with TEE/. These run in a protected environment and cost more. For example, on 11 October 2026 TEE/kimi-k3 was $3 and $15 per million tokens, against $2 and $10 for the standard version.
And as always, your chat app sees your messages too. Our data custody audit explains what the chat platforms themselves keep.
Should You Use NanoGPT?
NanoGPT fits you if:
- You want one key for many models, including community roleplay fine-tunes.
- You want to start with a very small deposit.
- You like the idea of a hard daily limit on a key.
- You play enough that a fixed $12 bill looks good.
Look elsewhere if:
- You want free API access. NanoGPT’s free model is website-only. See best free model for roleplay.
- You want the provider most guides and apps assume. That is still OpenRouter. See OpenRouter for roleplay.
- You want a short list of models that all run in a protected environment. See Chutes AI for roleplay.
For the three side by side, read OpenRouter vs NanoGPT vs Chutes.
Frequently Asked Questions
Is NanoGPT free?
Mostly no. NanoGPT’s pricing page lists one free model, and it works on the website only. Using NanoGPT with Janitor AI or SillyTavern needs an API key with a balance or a subscription. The smallest deposit is $0.10 with crypto or $1 with a card.
What does the NanoGPT subscription include?
On 11 October 2026 the Pro subscription costs $12 a month. It includes 60 million input tokens a week on a list of included models, 100 images a day, a 5% discount on eligible paid text models, and use through both the website and the API. The included list had 311 model entries and no Claude, GPT or Gemini models.
Is the NanoGPT subscription worth it for roleplay?
It depends on how much you play. At NanoGPT’s own pay-as-you-go prices, $12 buys about 2,000 half-hour sessions on DeepSeek V4 Flash, so a casual player spends less without the subscription. The subscription pays for itself when you play many sessions a day on mid-priced models or with a very large context.
What is the NanoGPT proxy URL for Janitor AI?
NanoGPT’s own Janitor AI guide gives https://api.nano-gpt.com/api/v1/chat/completions as the proxy URL. To show the model’s reasoning in the chat, it gives a second address: https://api.nano-gpt.com/api/v1thinking/chat/completions.
Does NanoGPT have Claude?
Yes, on pay-as-you-go. On 11 October 2026 NanoGPT listed Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and says it charges list price with no markup. Claude models are not in the subscription’s included list.
Why does Janitor AI show a proxy error with NanoGPT?
NanoGPT’s documentation gives three common causes. You chose a paid model that is not in your subscription and your balance is empty. The model ID is old, misspelled or unavailable, which usually returns a 404. Or the proxy URL or the API key is wrong, for example a key with an extra space.
Does NanoGPT store my chats?
NanoGPT’s privacy policy says that by default it does not store prompt or conversation content on its servers. It also says your prompt is passed to the company that runs the model, and that those companies may store and process it under their own policies.
Is NanoGPT the same as Karpathy’s nanoGPT?
No. nanoGPT is a code project by Andrej Karpathy for training small language models. NanoGPT at nano-gpt.com is a commercial service that sells access to AI models. The two share a name and nothing else.
The subscription price and contents, the deposit minimums and the “no markup” statement are from NanoGPT’s pricing page. Model prices, the count of 643 model entries and the 311-entry subscription list are from NanoGPT’s public model lists. The addresses, the Janitor AI steps, the key limits, the billing modes, the provider-selection rules and the error codes are from NanoGPT’s documentation. The privacy quotes are from its privacy policy, and the cancellation rule is from its refund policy, last updated 27 September 2026. All were read on 11 October 2026. The SillyTavern steps are from SillyTavern 1.19.0’s code. The session costs and the subscription arithmetic are our own calculations from those prices. We did not open a paid NanoGPT account or send requests, and nothing on this page is a first-hand test of speed or quality. Arcanum has no commercial relationship with NanoGPT.
NanoGPT, Janitor AI, SillyTavern, Chub, RisuAI, OpenRouter, Chutes, Claude, GPT, Gemini, DeepSeek, Kimi, GLM and Qwen are the property of their respective owners. Arcanum is not affiliated with any of them.