Best Model and API for Chub AI (2026): What Chub Users Actually Run, and What It Costs
The best model for Chub AI is DeepSeek V4.1 Flash, reached through an OpenRouter key, for most players who are willing to bring their own model. If you want one bill and no keys, the best choice is Chub’s own Soji, which comes with its $20 a month subscription. And if it has to cost nothing, the answer is a free model on OpenRouter or a Google key on Gemini’s free tier.
Chub AI is a front end. It builds a prompt from the character card, your chat and your settings, and sends it to a model somewhere else. So “which model” always comes with a second question, “through which API”, and this page answers both.
Most pages on this question give you one person’s favourite. This one starts from data: OpenRouter publishes which models Chub AI’s users actually run, in tokens, for the last 30 days. We read that page on 3 October 2026, matched every model to the price OpenRouter lists for it the same day, and read Chub’s subscription page, documentation and public model list.
A note on method. We have not run a controlled, model-by-model roleplay test of these models. OpenRouter says Chub’s users have run 348 of them, and nobody can test that honestly. The advice here comes from what Chub’s players use, what each model costs over a real session, and what Chub and the model makers publish. Where something is our own reading of the numbers, we say so.
If something is failing, not just undecided, start with Chub AI not working.
The Short Answer
| Your situation | Pick | Why |
|---|---|---|
| You want the best value and do not mind one key | DeepSeek V4.1 Flash on OpenRouter | Current DeepSeek model, 1M-token context, about $2.50 for 200 long turns before caching |
| You want no keys and a fixed price | Chub’s subscription (Soji) | $20 a month, unlimited messages for one user, nothing to set up |
| It has to be free, with an outside model | Nemotron 3 Ultra (free) on OpenRouter | The free model highest in Chub users’ list; daily request limits apply |
| It has to be free, from a big lab | Gemini Flash with a Google key | Google’s Flash models have a free API tier |
| You have a strong PC | A local model through Kobold or Ooba | The one connection Chub’s own documentation marks as free |
| You want the cheapest paid model that is still popular | DeepSeek V4 Flash, April version | About $0.23 for the same 200 turns |
| You love how Claude writes and accept the bill | Claude Sonnet 5.5 | About $17 for 200 turns; the older Sonnet 4.5 costs more |
The APIs Chub AI Supports
You choose the API in Chub’s Secrets screen, which Chub’s documentation calls “a necessary step before you can chat with a character”. Its documentation lists seven connections.
| Connection | What it reaches | Who pays whom |
|---|---|---|
| Mars (Chub’s own) | Chub’s hosted models, including Soji | $20 a month to Chub, in cryptocurrency |
| OpenRouter | Hundreds of models from many companies, one key | You pay OpenRouter per token; some models are free |
| OpenAI | GPT models, with a key or a custom address | You pay OpenAI per token |
| Anthropic | Claude models, with a key or a custom address | You pay Anthropic per token |
| Gemini models | You pay Google per token; Flash models have a free tier | |
| NovelAI | NovelAI’s writing models | Your NovelAI subscription |
| Kobold / Ooba | A model on your own computer, or a rented GPU | Free on your own hardware |
That list is from Chub’s documentation page on API connections, which was last updated about a year ago. Its examples of model names are old, but the list of connections matches what Chub’s users are doing today.
Which API should you use?
- OpenRouter, if you want choice. One key, and you can change model in a minute. Sixteen of the twenty models Chub’s users run most are from companies that have no connection of their own in Chub’s list, so OpenRouter is how people reach them.
- Chub’s own subscription, if you want no setup and no meter. More on it below.
- Google or Anthropic directly, if you only want that company’s models. You skip the middleman, and with Google you get the free tier.
- Kobold or Ooba, if the model is on your machine. Our local models guide covers what hardware that takes.
One warning, in Chub’s own words. Its documentation separates official keys from “reverse proxies”, meaning addresses run by third parties that promise free access:
“Reverse proxies, on the other hand, are created by unauthorized third parties who gain access to your data and IP address.”
It recommends official keys, and so do we. A free proxy can read your chats, and it can disappear overnight.
What Chub AI Users Actually Run
OpenRouter publishes a usage page for every app that connects to it. On 3 October 2026, the page for Chub AI showed 123B total tokens, 348 models used, activity since September 2024, and a rank of #7 in Roleplay.
It also lists the models Chub’s users ran the most in the last 30 days. Here is that list with the price OpenRouter shows for each model on the same day, in US dollars per million tokens.
| # | Model | Tokens, last 30 days | Input | Output | Context window |
|---|---|---|---|---|---|
| 1 | DeepSeek V3.2 | 14.3B | $0.28 | $0.42 | 164K |
| 2 | GLM 5.2 (Z.ai) | 10.9B | $0.41 | $3.99 | 1M |
| 3 | DeepSeek V4.1 Flash | 7.97B | $0.30 | $1.20 | 1M |
| 4 | Claude Sonnet 4.5 (Anthropic) | 7.09B | $3.00 | $15.00 | 1M |
| 5 | DeepSeek V4 Pro 0813 | 7.03B | $0.22 | $4.20 | 1M |
| 6 | DeepSeek V4 Flash, April version | 6.85B | $0.03 | $0.06 | 1M |
| 7 | Nemotron 3 Ultra (NVIDIA) | 5.97B | free version: $0 | $0 | 1M (free version) |
| 8 | DeepSeek V4 Pro, April version | 5.47B | $0.21 | $0.42 | 1M |
| 9 | DeepSeek V3 0324 | 3.26B | $0.29 | $1.14 | 164K |
| 10 | DeepSeek V3.1 | 3.22B | $0.25 | $0.95 | 164K |
| 11 | DeepSeek V3.1 Terminus | 3.15B | $0.27 | $1.00 | 164K |
| 12 | DeepSeek V4 Flash 0731 | 3.15B | $0.02 | $1.28 | 1M |
| 13 | Claude Opus 4.5 (Anthropic) | 2.88B | $5.00 | $25.00 | 200K |
| 14 | GLM 5.3 Flash (Z.ai) | 2.79B | $0.15 | $0.50 | 1M |
| 15 | MiniMax M3 | 2.43B | $0.30 | $1.20 | 1M |
| 16 | Gemma 4 31B (Google) | 2.38B | $0.09 (a free version exists) | $0.34 | 262K |
| 17 | DeepSeek V3.2 Exp | 2.34B | $0.27 | $0.41 | 164K |
| 18 | MiMo-V2.5-Pro (Xiaomi) | 2.26B | $0.44 | $0.87 | 1M |
| 19 | GLM 5 (Z.ai) | 2.10B | $0.60 | $1.92 | 205K |
| 20 | Gemini 3.8 Flash (Google) | 1.61B | $0.75 | $3.75 | 1M |
Before reading anything into it, know what this list is and is not.
- It only counts OpenRouter. It does not include Chub’s own models, people who connect straight to Anthropic, Google or OpenAI, or anyone running a local model. Chub’s subscribers are invisible here.
- It counts tokens, not people. One player with a huge context setting can outweigh twenty casual ones.
- It does not say whether a model was run in its free or paid version. For Nemotron 3 Ultra we show the free version, because it has one. That is our assumption about why it ranks so high, not something the page states.
- The match between a name on the usage page and a price on the model list is ours.
- Prices on OpenRouter move. These are the prices on 3 October 2026.
With that said, four things stand out.
1. Chub on OpenRouter is a DeepSeek platform. Ten of the twenty models are DeepSeek. Together they are about 56.7 billion of the roughly 97 billion tokens in the top 20, which is close to 58%. That is our sum of OpenRouter’s figures.
2. The most used model is ten months old. DeepSeek V3.2 was listed on OpenRouter on 1 December 2025. Its context window is about one sixth the size of the current DeepSeek models. It is still first. Presets and habit keep people on a model long after a newer one arrives, and V3.2 is cheap on output.
3. Chub’s users pay for Claude, and for old Claude. Claude Sonnet 4.5 is fourth and Claude Opus 4.5 is thirteenth. Together they are about 10% of the top 20’s tokens. On Janitor AI’s page, which we read the day before, no Claude model was in the top 20 at all. And both of these are 2025 models. Sonnet 4.5 costs more than its successor: $3 and $15 per million tokens, against $2 and $10 for Claude Sonnet 5.5.
4. Nearly everything else costs under $0.50 per million input tokens. Input price is what matters in roleplay, because every turn sends the whole conversation again.
What a Session Costs on Each Model
A price per million tokens is hard to feel. So here is one long session, priced on the models Chub’s users choose.
The turn: the model receives the character card, your persona, lorebook entries and the recent conversation — 40,000 tokens in total — and writes a 500-token reply. It is the same turn we use on our DeepSeek, Gemini and Claude pages, so the numbers compare.
| Model | 200 turns |
|---|---|
| Nemotron 3 Ultra, free version | $0 (within the daily request limit) |
| DeepSeek V4 Flash, April version | $0.23 |
| DeepSeek V4 Flash 0731 | $0.28 |
| Gemma 4 31B | $0.75 ($0 on its free version) |
| GLM 5.3 Flash | $1.25 |
| DeepSeek V4 Pro, April version | $1.71 |
| DeepSeek V4 Pro 0813 | $2.18 |
| DeepSeek V3.2 | $2.28 |
| DeepSeek V4.1 Flash | $2.52 |
| MiniMax M3 | $2.52 |
| GLM 5.2 | $3.68 |
| Gemini 3.8 Flash | $6.38 |
| Claude Sonnet 5.5 | $17.00 |
| Claude Sonnet 4.5 | $25.50 |
| Claude Opus 4.5 | $42.50 |
These numbers are our arithmetic on OpenRouter’s listed prices, and they are a worst case: a full 40,000-token context on every turn, no caching discount and no thinking.
- Caching makes it cheaper. Many providers charge far less for input they have already seen.
- Thinking makes it more expensive. Several of these models reason before they answer, and the reasoning is billed as output.
- A smaller context makes it much cheaper. At 16,000 tokens of context, every number above is less than half.
Now put Chub’s own subscription beside that table. $20 a month is about eight of those 200-turn sessions on DeepSeek V4.1 Flash, at the worst-case price. If you play less than that, your own key is cheaper. If you play every day, or you would otherwise be on Claude, the flat fee wins quickly. That comparison is only about money; it says nothing about which model writes better.
Chub’s Own Models: Soji and the $20 Subscription
Chub’s documentation calls its hosted models Mars, “our suite of uncensored language models finetuned on public datasets as well as our own proprietary data”. It adds that this data “does NOT include public or private chats taken from Chub AI”.
The flagship is Soji. Chub’s patch notes introduced it as “a 671B-parameter finetune of DeepSeek V3 0324, at 60K context for Mars subscribers”. Chub’s subscription page on 3 October 2026 lists it as “Soji 1.6T 60K” and does not explain the label further.
There is something worth noticing in that. Chub built its flagship on a DeepSeek model, and Chub’s users who bring their own model mostly choose DeepSeek too. The two halves of the platform have arrived at the same family.
What the subscription page says $20 a month buys:
- “Unlimited access to all LLMs, including Soji”
- Unlimited voice and multimedia generation
- Early access to new models, and access to future ones
- An API that “can be used anywhere that supports OpenAI’s APIs”, which means you can use Chub’s models from another front end such as SillyTavern
And the conditions:
- “Unlimited” means one user. Chub’s page: “Unlimited for a single user, similar to Netflix. Concurrent requests from the same account will return a 429 error.” Two chats generating at once on one account will fail.
- Payment is in cryptocurrency only. There is no card form. Chub says this keeps it free of pressure from payment companies.
- The trial is one-time. “One-time free trial credits amounting to a few dozen messages are sent out to new users regularly.”
On 3 October 2026 Chub’s public API listed these model names: soji, asha, mythomax, mixtral, mistral, mobile and testing. Chub’s public pages describe only Soji. People search for a “free mobile model” on Chub; we could not confirm from any page we could read what the mobile model is or whether it is free, so we will not say.
Is Soji better than DeepSeek through your own key? We have not compared them, and we will not guess. What we can say is what differs on paper: Soji has a 60K context and a fixed monthly price; current DeepSeek models on OpenRouter have a 1M context and a price per token. If you switch to Soji, accept the preset Chub offers. Its patch notes say Soji “has different ideal settings than other models”.
The Best Paid Pick: DeepSeek V4.1 Flash on OpenRouter
DeepSeek V4.1 Flash is our default for a paid model on Chub, for the same four reasons as on any front end.
- It is the current model. It was released on 10 September 2026 and it is the one DeepSeek is improving.
- It has a 1M-token context window. DeepSeek V3.2, the popular one, stops at about 164,000.
- It is cheap on input, where roleplay spends its money: $0.30 per million tokens.
- It is already third in Chub users’ list after three weeks.
The model name on OpenRouter is deepseek/deepseek-v4.1-flash.
Should you stay on DeepSeek V3.2? If you already use it and like how it writes, there is no urgent reason to move. Many presets are tuned for it. Know what you give up: a much smaller context window, and a model that DeepSeek itself no longer serves on its own API.
Can you connect to DeepSeek’s own API from Chub? DeepSeek has no entry of its own in Chub’s documented list. DeepSeek’s API uses the same format as OpenAI’s, and Chub’s documentation says the OpenAI connection accepts a custom address as well as a key, so that is the route people describe. We have not set that up on Chub ourselves, so we cannot give you tested steps. If you try it, the model name on DeepSeek’s own API is deepseek-flash; the older names deepseek-chat and deepseek-reasoner were retired on 24 July 2026. Through OpenRouter none of this matters.
The choice between DeepSeek’s own models is covered in the best DeepSeek model for roleplay.
The Best Free Options
Chub’s own models stop being free when the trial credits run out. After that there are three honest free routes.
A free model on OpenRouter
On 3 October 2026 OpenRouter listed 17 models with a free version. The one highest in Chub users’ list is NVIDIA’s Nemotron 3 Ultra:
nvidia/nemotron-3-ultra-550b-a55b:free
Gemma 4 31B from Google is also in Chub’s top 20 and also has a free version: google/gemma-4-31b-it:free.
Free on OpenRouter comes with rules:
- 20 requests a minute, for everyone.
- 50 requests a day if you have bought fewer than 10 credits in total.
- 1,000 requests a day once you have bought at least 10 credits, ever.
- A negative balance blocks free models too.
- Every reroll is a request.
There is no free DeepSeek on OpenRouter today. None of the 17 free models is a DeepSeek model. Older guides for Chub tell you to type a DeepSeek name ending in :free. That is why those setups return “model not found”. Our guide to the best free model for roleplay follows the whole free list.
A Google key
Google’s Gemini API has a free tier for its Flash models, and Chub has a Google connection. Two things to know: Google does not publish the free tier’s rate limits outside its own console, and its terms allow free-tier content to be used to improve its products. Details are in the best Gemini model for roleplay.
A model on your own computer
Chub’s documentation marks one connection as free: Kobold and Ooba, “free if being hosted from your machine with the only limitations being hardware requirements”. Chub is a website, so your local model must be reachable from your browser; follow the setup in Chub’s own documentation for that.
Why Chub’s Users Still Pick Claude
Claude is the expensive choice everywhere, and on Chub it is clearly a popular one. Two practical facts explain some of what the list shows.
The newer Claude is cheaper. Claude Sonnet 5.5 is $2 input and $10 output per million tokens on OpenRouter. Sonnet 4.5, the one in Chub users’ top five, is $3 and $15. If you are on Sonnet 4.5 only because an old guide named it, you are paying half as much again for the older model.
But the newer Claude is stricter about settings. Anthropic’s documentation says that from Claude Opus 4.7 onward, setting temperature, Top P or Top K to a non-default value returns an error. Sonnet 5.5 also always thinks before it answers on OpenRouter’s list. The 4.5 models accept the old sampler settings and do not have to think. So a preset written for Sonnet 4.5 can fail when you point it at Sonnet 5.5. We do not know that this is why people stay; it is one concrete difference you will meet if you switch. Reset the sampler settings to default first. More in the best Claude model for roleplay.
Presets, Thinking and Empty Replies
Three things go wrong when people change model on Chub, and none of them is the model’s quality.
1. The old preset comes along. Chub’s documentation says presets “are model-specific”. Its patch notes say that when you switch model, “some will now ask if you want to pull in the settings of the most popular preset for that model”. Say yes. A preset tuned for one model is a common reason the next model writes badly.
2. The model thinks, and the reply comes back empty. On OpenRouter’s list, thinking is on by default for DeepSeek V4.1 Flash, GLM 5.2 and Nemotron 3 Ultra, and always on for GLM 5.3 Flash and Gemini 3.8 Flash. Thinking counts as output. If your maximum tokens setting is small, the model can spend all of it thinking and send back nothing. That is the usual cause of “empty response from API”. OpenRouter’s documentation says that retrying does not help and that you should raise the limit or reduce the reasoning.
3. The temperature does nothing, or breaks everything. DeepSeek’s own API ignores temperature while thinking is on. Google has deprecated temperature, Top P and Top K on current Gemini models. Current Claude models reject non-default values. Chub’s own rule of thumb for the penalty settings is still useful where they apply:
“If it’s being repetitive, the repetition, presence, and/or frequency penalties are too low. If it’s dropping common words like ‘the’, ‘an’, and pronouns, those penalties are too high.”
Our Recommendation
- Use the trial credits to see whether you like Soji. They are one-time, so spend them on a real scene.
- If you want no setup and you play a lot, subscribe. $20 is about eight long sessions on a cheap model, and a fraction of one month on Claude.
- If you want choice, get an OpenRouter key and start on DeepSeek V4.1 Flash.
- If it must be free, use Nemotron 3 Ultra’s free version or a Google key, and expect limits.
- If DeepSeek’s style is not for you, try what Chub’s users try next: GLM 5.2, MiniMax M3, MiMo-V2.5-Pro. They cost about the same.
- Pick Claude for its writing, and pick Sonnet 5.5 over 4.5 unless you have a reason to stay.
- Accept the popular preset when you switch model, and leave the maximum tokens generous on any model that thinks.
- Use official keys, not free proxies.
To compare model families for roleplay in general, see the best LLM for roleplay. For what we think of the platform itself, see our Chub AI review. If you are new to the site and want the menus explained, start with how to use Chub AI.
Frequently Asked Questions
What is the best model for Chub AI?
For most people, DeepSeek V4.1 Flash through an OpenRouter key. It has a 1M-token context window and costs $0.30 input and $1.20 output per million tokens, so a long session costs a couple of dollars at most. If you would rather pay one fixed bill and manage no keys, Chub’s own Soji model is included in its $20 a month subscription. If it has to be free, use a free model on OpenRouter such as NVIDIA’s Nemotron 3 Ultra, or a Google key with Gemini’s free tier.
Which API should I use on Chub AI?
OpenRouter is the most flexible choice: one key reaches hundreds of models, including the DeepSeek, GLM, Claude and Gemini models Chub users run most. Use Chub’s own subscription if you want unlimited messages for a fixed price with no setup. Use a Google or Anthropic key directly if you only want that one company’s models. Use Kobold or Ooba if you run a model on your own computer.
Is there a free model on Chub AI?
Chub’s own models are not free after the trial: new users get one-time trial credits worth a few dozen messages. After that, free play means bringing your own model. The options are a free model on OpenRouter, which is limited to 50 requests a day or 1,000 a day once you have bought 10 credits in total; a Google key, because Gemini’s Flash models have a free API tier; or a model running on your own computer through Kobold or Ooba.
What is Soji on Chub AI?
Soji is Chub’s flagship hosted model. Chub’s patch notes introduced it as a 671-billion-parameter finetune of DeepSeek V3 0324 with a 60K context, for subscribers. Its subscription page on 3 October 2026 lists it as ‘Soji 1.6T 60K’ and says the $20 a month subscription gives unlimited access to it and to Chub’s other models.
What is the difference between Mars and Soji on Chub AI?
Mars is the name Chub’s documentation uses for its subscription and its suite of hosted models. Soji is one model inside it, the flagship. On 3 October 2026 Chub’s public API listed these model names: soji, asha, mythomax, mixtral, mistral, mobile and testing.
What is the best DeepSeek model for Chub AI?
DeepSeek V4.1 Flash, which on OpenRouter has the model name deepseek/deepseek-v4.1-flash. The model Chub users run most is the older DeepSeek V3.2, which costs less for output but has a context window of about 164,000 tokens instead of 1M. Both are cheap enough that a long session costs around two dollars before any caching discount.
Is Claude a good model for Chub AI?
Chub users choose it more than users of similar sites do: Claude Sonnet 4.5 was the fourth most used model on Chub’s OpenRouter page. It is expensive. The 200-turn session that costs about $2.50 on DeepSeek V4.1 Flash costs about $25.50 on Claude Sonnet 4.5. The newer Claude Sonnet 5.5 is cheaper, at about $17 for the same session, but it rejects requests that set temperature, Top P or Top K to a non-default value.
Why does Chub AI say empty response from API?
Usually because the model thinks before it answers and spent its whole token limit on thinking. OpenRouter’s documentation says retrying does not help, and that the fix is to raise the maximum tokens or limit the reasoning. Many of the models Chub users run think by default, so this is a common result of switching to a newer model with an old preset.
The usage figures are from OpenRouter’s public app page for Chub AI (“Top models used by Chub AI this month”, last 30 days), read on 3 October 2026. Prices, context windows, free models and thinking defaults are from OpenRouter’s model list on the same day; the free-model request limits and the note on empty responses are from OpenRouter’s documentation. The subscription price and conditions are from Chub’s subscription page, the connection list and the reverse-proxy warning from its documentation, the Soji description and the preset advice from its patch notes, and the model names from its public API endpoint, all read on 3 October 2026. The DeepSeek share, the match between usage names and price-list entries, and every session cost are our own arithmetic. We have not run a controlled comparison of how these models write, and we have not tested Chub’s own models against outside ones.
Product names and trademarks are the property of their respective owners. Arcanum is not affiliated with Chub AI, OpenRouter, DeepSeek, Anthropic, Google, NVIDIA or any model maker named here.