Best Model for Janitor AI (2026): What Players Actually Use, and What to Pick
The best model for Janitor AI is DeepSeek V4.1 Flash through a proxy, for most players who are willing to do ten minutes of setup. If you want no setup, the best model is the one you already have: Janitor’s own JLLM, which is free and has no message limit. And if you want an outside model but cannot pay anything, the answer today is NVIDIA’s Nemotron 3 Ultra on OpenRouter’s free tier.
Most pages that answer this question give you one person’s favourite. This one starts from something better: OpenRouter publishes which models Janitor AI’s users actually run, in tokens, for the last 30 days. So we can show you what thousands of players have settled on, what each of those models costs, and where the popular choice is not the best one.
Everything below was checked on 2 October 2026: the usage figures on OpenRouter’s public page for Janitor AI, the prices and model names on OpenRouter’s model list, and Janitor’s own help centre.
A note on method. We have not run a controlled, model-by-model roleplay test of these models. Nobody can honestly test 337 models, and that is how many OpenRouter says Janitor’s users have run. The advice here comes from what players actually use, what each model costs over a real session, and what Janitor and the model makers publish. Where something is our own reading of the numbers, we say so.
If you have never connected an outside model, read what a Janitor AI proxy is first. This page only answers which model.
The Short Answer
| Your situation | Pick | Why |
|---|---|---|
| No setup, no money | JLLM (Janitor’s own model) | Free, unlimited, already switched on |
| No money, but you want an outside model | Nemotron 3 Ultra (free) on OpenRouter | The free model Janitor users run the most; daily request limits apply |
| A few cents an evening is fine | DeepSeek V4.1 Flash | Newest DeepSeek, 1M-token context, about $2.50 for 200 long turns before caching |
| The cheapest paid model that is still popular | DeepSeek V4 Flash (the April version) | About $0.23 for the same 200 turns |
| You like JLLM but your long chats forget | janitor+ | Five times the context on Janitor’s own model |
| You want outside models with no proxy setup | janitor+ Router | Pick from a list, pay per token from a wallet |
| You love how Claude writes and accept the bill | Claude Sonnet 5.5 | About $17 for the same 200 turns before caching |
The rest of the page explains each row.
The Three Ways to Get a Model on Janitor AI
“Which model” depends on how you connect, so here is the map first.
1. JLLM, Janitor’s own model. It is the default. Janitor’s help centre describes it as “free, built-in, and ready to go out of the box” and calls it the “house brand”. On the free plan it works with a context limit of around 9,000 tokens. A second version, jllm-short, writes shorter and faster-paced replies and has been free for everyone since 15 July 2026.
2. A proxy. You open an account with a model provider, copy an API key, and paste three things into Janitor’s API settings: a URL, the key and a model name. Janitor charges nothing for this, on any plan. You pay the provider. Our API key guide and proxy guide cover the steps.
3. The janitor+ Router. This is a proxy that lives inside Janitor. You pick a model from a list in your settings and pay per token from a wallet. It needs the janitor+ subscription ($12.99 a month) and it is web only for now. Janitor AI pricing has the details.
Most of this page is about route 2, because that is where the choice is widest and where most of the searching happens.
What Janitor AI Users Actually Run
OpenRouter is a service that puts hundreds of models behind one API key, and it is the most common proxy route for Janitor. It also publishes usage figures for the apps that connect to it. Janitor AI has its own page there. On 2 October 2026 that page said Janitor’s users had run 337 different models, and that Janitor ranked #2 in the Roleplay category.
It also lists the models Janitor’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. Prices are in US dollars per million tokens.
| # | Model | Tokens, last 30 days | Input | Output | Context window |
|---|---|---|---|---|---|
| 1 | DeepSeek V3.2 | 245B | $0.28 | $0.42 | 164K |
| 2 | Nemotron 3 Ultra (NVIDIA) | 96B | free version: $0 | $0 | 1M (free version) |
| 3 | GLM 5.2 (Z.ai) | 65.2B | $0.41 | $3.99 | 1M |
| 4 | DeepSeek V4 Pro, April version | 52.7B | $0.21 | $0.42 | 1M |
| 5 | DeepSeek V4 Flash, April version | 48.5B | $0.03 | $0.06 | 1M |
| 6 | MiniMax M3 | 26.5B | $0.30 | $1.20 | 1M |
| 7 | GLM 5.3 Flash (Z.ai) | 26.2B | $0.15 | $0.50 | 1M |
| 8 | DeepSeek V3 0324 | 23.8B | $0.29 | $1.14 | 164K |
| 9 | DeepSeek V4.1 Flash | 21.2B | $0.30 | $1.20 | 1M |
| 10 | MiMo V2.5 Pro (Xiaomi) | 21B | $0.44 | $0.87 | 1M |
| 11 | DeepSeek V3.1 | 21B | $0.25 | $0.95 | 164K |
| 12 | DeepSeek V4 Flash 0731 | 19.4B | $0.005 | $1.28 | 1M |
| 13 | DeepSeek R1 0528 | 17.3B | $0.50 | $2.15 | 164K |
| 14 | DeepSeek V4 Pro 0813 | 17B | $0.66 | $1.98 | 1M |
| 15 | GLM 4.7 (Z.ai) | 14.1B | $0.60 | $2.20 | 205K |
| 16 | Gemini 3.8 Flash (Google) | 13.9B | $0.75 | $3.75 | 1M |
| 17 | Kimi K2.5 (Moonshot) | 13.8B | $0.45 | $2.25 | 262K |
| 18 | GLM 5 (Z.ai) | 13.2B | $0.60 | $1.92 | 205K |
| 19 | Nemotron 3 Super (NVIDIA) | 12.3B | free version: $0 | $0 | 262K |
| 20 | Kimi K2.7 Code (Moonshot) | 11.4B | $0.67 | $3.35 | 262K |
Before reading anything into it, know what this list is and is not.
- It only counts OpenRouter. It does not include JLLM, the janitor+ Router, people who connect to DeepSeek’s own API, or other hosts. JLLM is the default for every new account, so it is almost certainly the most used model on Janitor overall. Nobody publishes that number.
- It counts tokens, not people. One player with a huge context setting can outweigh twenty casual ones.
- The usage page does not say whether a model was run in its free or paid version. For the two NVIDIA models we show the free version’s price, because both have one. That is our assumption about why they rank 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. OpenRouter’s usage page shows display names; we matched each to the model with that name and date on the price list.
With that said, five things stand out.
1. Janitor on OpenRouter is a DeepSeek platform. Nine of the twenty models are DeepSeek. Together they are about 466 billion of the roughly 780 billion tokens in the top 20, which is close to 60%. 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. It has a context window about one sixth the size of the current DeepSeek models. It is still first, with more tokens than the next three combined. Presets, guides and habit keep people on a model long after a newer one arrives. V3.2 is also cheap on output, at $0.42 per million tokens.
3. The second most used model is free. That is the clearest signal of what “best free model for Janitor AI” means in practice today.
4. No Claude and no GPT model is in the top 20. The only model from a large Western lab is Gemini 3.8 Flash, at number 16. Search results for this question are full of Claude and ChatGPT advice. Janitor’s actual users, on this route, have gone somewhere else.
5. Everything in the top 15 costs under $1 per million input tokens. Input price is what matters in roleplay, because every turn sends the whole conversation again. Players have found that out with their wallets.
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 people actually choose.
The turn: the model receives your bot’s definition, your persona, chat memory and the recent conversation — 40,000 tokens in total — and writes a 500-token reply. This is the same turn we use on our DeepSeek, Gemini and Claude model pages, so the numbers compare.
| Model | One turn | 200 turns |
|---|---|---|
| Nemotron 3 Ultra, free version | $0 | $0 (within the daily request limit) |
| DeepSeek V4 Flash, April version | $0.0012 | $0.23 |
| GLM 5.3 Flash | $0.0063 | $1.25 |
| DeepSeek V3.2 | $0.0114 | $2.28 |
| DeepSeek V4.1 Flash | $0.0126 | $2.52 |
| MiniMax M3 | $0.0126 | $2.52 |
| MiMo V2.5 Pro | $0.0178 | $3.57 |
| GLM 5.2 | $0.0184 | $3.68 |
| Gemini 3.8 Flash | $0.0319 | $6.38 |
| Claude Sonnet 5.5 | $0.0850 | $17.00 |
These numbers are our arithmetic on OpenRouter’s listed prices, and they are a worst case. They assume no caching and no thinking.
- Caching makes it cheaper. Many providers charge far less for input they have already seen. On DeepSeek’s own API, the same 200 turns on V4.1 Flash come to about $0.29 with cache hits. Whether you get that discount through a router depends on which company hosts the model.
- Thinking makes it more expensive. Several of these models reason before they answer, and the reasoning is billed as output. More on that below.
- A smaller context makes it much cheaper. 40,000 tokens is a long campaign. If your context setting is 16,000 tokens, every number above is less than half.
The reading is simple. On the models in Janitor users’ top ten, a long evening costs between nothing and about four dollars. On Claude it costs more than four times the most expensive of those. That gap, more than any difference in writing, is why the top 20 looks the way it does.
The Best Paid Pick: DeepSeek V4.1 Flash
DeepSeek V4.1 Flash is our default for a paid proxy model, for four reasons that all point the same way.
- It is the current model. It was released on 10 September 2026 and it is the model 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 in Janitor users’ top ten after three weeks.
The model names to type:
| Where you connect | Model name |
|---|---|
| DeepSeek’s own API | deepseek-flash |
| OpenRouter | deepseek/deepseek-v4.1-flash |
If your setup still says deepseek-chat or deepseek-reasoner and you connect to DeepSeek directly, that is why it stopped working: DeepSeek retired both names on 24 July 2026. The full step-by-step is in how to use DeepSeek on Janitor AI, and the choice between DeepSeek’s own models is in the best DeepSeek model for roleplay.
Should you stay on DeepSeek V3.2 instead? If you already use it and like how it writes, there is no urgent reason to move. It is slightly cheaper per session and 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. It stays available only because other companies host the open weights.
What about the very cheap April version of V4 Flash? At $0.03 input it is the cheapest popular model on the list by a wide margin, and it has the 1M window. DeepSeek has replaced it twice since April, so treat it as a bargain that may not stay listed. For a campaign you plan to run for months, pick the current model.
The Best Free Options
There are two kinds of free on Janitor, and they are very different.
JLLM: free with no limits
Janitor’s own model is the only free option with no daily cap and no setup. The free plan gives it a context of around 9,000 tokens. That is small: your bot’s definition, your persona and your chat memory all have to fit in it before any conversation does. Janitor’s own guide to tokens advises keeping the permanent part of a bot under about 1,500 tokens for that reason.
If you like how JLLM writes and the only problem is that long chats forget, the fix is not a different model. It is janitor+, which gives the same model five times the context. We explain the memory side in Janitor AI chat memory.
A free outside model: Nemotron 3 Ultra
If you want an outside model at zero cost, the one Janitor users have chosen is NVIDIA’s Nemotron 3 Ultra. On OpenRouter its free version has the model name:
nvidia/nemotron-3-ultra-550b-a55b:free
OpenRouter lists it with a 1M-token context window. It thinks before it answers by default, so replies start more slowly than on JLLM.
Free on OpenRouter comes with rules, and they explain most “it suddenly stopped working” moments:
- 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. The credits do not need to stay in your account.
- A negative balance blocks free models too. OpenRouter’s documentation says so directly.
- Every reroll is a request. Fifty a day goes quickly if you swipe a lot.
There is also a privacy cost. Janitor’s own troubleshooting article tells users to switch “Model Training” on in their OpenRouter settings to use free models. In plain words: free models are often free because the host may use what you send. If that matters to you, pay the few cents.
The free DeepSeek route is closed
For most of 2025, “free model for Janitor AI” meant free DeepSeek on OpenRouter. Janitor’s own quickstart proxy guide still gives that setup, with the model name deepseek/deepseek-chat-v3-0324:free. That article was last updated in August 2025.
On 2 October 2026 OpenRouter listed 17 models with a free version, and none of them was a DeepSeek model. The same name without :free still works, as a paid model at $0.29 input and $1.14 output per million tokens. So if you followed an older guide and get a “model not found” error, nothing is wrong with your key. The free version of that model is gone.
Free lists change every few weeks. Our guide to the best free model for roleplay follows the whole free list, not only the Janitor side of it.
Quick Answers by Model Family
Many people search for the best model of one brand for Janitor. The honest answer is the same for all of them: Janitor does not change which model in a family is best. It only changes how you reach it. So here is each family in a few lines.
DeepSeek. V4.1 Flash, as above. Full comparison: best DeepSeek model for roleplay.
Gemini. Gemini 3.8 Flash is the one Janitor users run, and the only model from a big Western lab in their top 20. It costs $0.75 input and $3.75 output per million tokens, and Google’s API has a free tier for its Flash models. Details and the free-tier conditions are in best Gemini model for roleplay.
Claude. Claude Sonnet 5.5 for most uses, at $2 input and $10 output. It is the expensive choice on Janitor, and no Claude model is in the top 20. One practical warning: current Claude models reject requests that set temperature or the other sampling values to anything but the default, so an old preset can make every message fail. See best Claude model for roleplay.
GPT. OpenRouter’s list on 2 October shows GPT-6.1 Sol at $2 input and $10 output, and the small GPT-6 Luna at $0.10 and $0.50. No GPT model is in Janitor users’ top 20. We have not written a model-by-model page for OpenAI yet, so we will not rank them here.
Grok. Grok 4.7 is listed at $2 input and $6 output with a 500K window. Not in the top 20 either.
GLM (Z.ai). The second most popular family on Janitor after DeepSeek, with four models in the top 20. GLM 5.3 Flash is the cheap one at $0.15 and $0.50. Note that on OpenRouter the GLM 5.3 models always think before answering; it cannot be turned off there.
Kimi, MiniMax, MiMo. All three have a model in the top 20, all priced under $0.70 per million input tokens. They are reasonable things to try if DeepSeek’s style does not suit you. We have not compared their writing.
Mistral and Qwen. People search for both, and neither is in the top 20. Qwen has one free model on OpenRouter today (qwen/qwen3.8-27b:free). Mistral has none.
A local model (KoboldCpp, Ollama). Janitor is a website, so it can only use a model it can reach over the internet. A model on your own PC needs a public address before Janitor can talk to it, and that is a real security step, not a checkbox. If you want local models, a frontend that runs on your own machine is the easier home for them. Start with local models for AI RPGs.
Thinking Models: Why Replies Are Slow, or Show Their Reasoning
Many 2026 models think before they answer: they write out their reasoning first, then the reply. On OpenRouter’s list, thinking is on by default for DeepSeek V4.1 Flash, Nemotron 3 Ultra, GLM 5.2 and Kimi K2.5, and it is always on for the GLM 5.3 models and DeepSeek R1.
On Janitor this shows up in three ways.
- Replies start late. The model is reasoning before the first word of the reply.
- You pay for the reasoning. It is billed as output, which is the expensive side.
- Sometimes the reasoning lands in the chat. Janitor’s help article on one provider warns that its direct API “does not hide reasoning”, and suggests a model that does not think if you want “a cleaner experience”.
There is one more effect that confuses people. On DeepSeek, when thinking is on, the model ignores your temperature setting and gives no error. The slider moves and nothing changes. We explain that in the DeepSeek model guide.
So for fast, chatty scenes, choose a model where thinking is optional and turn it off if your provider’s settings allow it. For a campaign with rules, stats and a lot of state — the kind we describe in how to run an RPG on Janitor AI — thinking is worth the wait.
Temperature and Context: Two Settings That Matter More Than the Model
Temperature. Janitor’s own guide suggests a starting point: “Start at 0.6”, raise it by 0.1 if replies are bland, and lower it if the model starts producing nonsense. It adds that you should “keep the temperature below 1.0” unless you want chaos, and that models differ: “what’s mild in Claude might be extra spicy in Mistral”. That last line is the important one. A temperature you copied from a preset for one model is not a setting for another.
Context size. A 1M-token window does not mean you should send a million tokens. Every turn sends your whole context again, and you pay for all of it. Most of the cost table above is context, not reply. Set the context to what the story needs. If the bot forgets things, fix that with chat memory first, and with a bigger window second.
Every slider in the panel, with its real range and what to set per model, is in Janitor AI generation settings.
Do You Need janitor+ for a Better Model?
No. This is the most common misunderstanding, so it is worth saying plainly.
- Your own proxy is free on Janitor’s side, on both plans. janitor+ does not make a proxy model better or cheaper.
- janitor+ improves Janitor’s own model. Five times the context, priority at busy times, stronger rerolls.
- The janitor+ Router is convenience. No key, no URL, a list with prices. Janitor does not publish the Router’s model list outside your settings, so check that the model you want is there before you subscribe for it. Subscribers get 10% off some popular models and a $5 starting credit.
Janitor itself does not present the Router as the upgrade path. Its help article says: “For everyday roleplay and conversation, JLLM+ is still the recommended chat model. Router exists to give you additional options.”
Our Recommendation
- Start on JLLM. It is free and unlimited, and you learn what you want from a model before paying for one.
- When you want better writing or longer memory, connect DeepSeek V4.1 Flash. Directly, with the name
deepseek-flash, or through OpenRouter asdeepseek/deepseek-v4.1-flash. - If it must be free, use Nemotron 3 Ultra’s free version on OpenRouter, and expect daily limits.
- If DeepSeek’s style is not for you, try the next families Janitor users chose: GLM, MiniMax, MiMo, Kimi. They cost about the same.
- Pick Claude or Gemini only for their writing, knowing a session costs several times more.
- Whatever you pick, check the model name on the provider’s list today. Most broken setups are an old name from an old guide.
If a connection fails after all that, Janitor AI not working lists the errors and what each one means. And if you want to compare model families for roleplay in general, not only on Janitor, see the best LLM for roleplay.
Frequently Asked Questions
What is the best model for Janitor AI?
For most players who are willing to set up a proxy, DeepSeek V4.1 Flash. It has a 1M-token context window and costs $0.30 input and $1.20 output per million tokens, so a long evening is a few cents to a couple of dollars. If you want no setup at all, stay on Janitor’s own model, JLLM, which is free and unlimited. If it has to be an outside model and it has to be free, NVIDIA’s Nemotron 3 Ultra is the free model Janitor users run the most on OpenRouter.
What is the best free model for Janitor AI?
Janitor’s own JLLM is the only free option with no daily limit and no setup. Among outside models, the free one Janitor users run most on OpenRouter is NVIDIA’s Nemotron 3 Ultra, with the model name nvidia/nemotron-3-ultra-550b-a55b:free. OpenRouter limits free models to 20 requests a minute and 50 requests a day, or 1,000 a day once you have bought at least 10 credits in total.
Which model do most Janitor AI users use?
On OpenRouter, DeepSeek by a wide margin. On 2 October 2026, OpenRouter’s public page for Janitor AI listed nine DeepSeek models among the 20 most used, and together they made up about 60% of the tokens in that top 20. The single most used model was DeepSeek V3.2. This only counts traffic that goes through OpenRouter, so it does not include Janitor’s own JLLM, the janitor+ Router or people who connect to DeepSeek directly.
What is the best DeepSeek model for Janitor AI?
DeepSeek V4.1 Flash. On DeepSeek’s own API the model name is deepseek-flash. On OpenRouter it is deepseek/deepseek-v4.1-flash. The older names deepseek-chat and deepseek-reasoner were retired on 24 July 2026 and no longer work on DeepSeek’s API. Many Janitor users still run the older DeepSeek V3.2 through OpenRouter, which costs less for output but has a context window of about 164,000 tokens instead of 1M.
Is the free DeepSeek model from Janitor’s proxy guide still available?
No. Janitor’s quickstart proxy guide, last updated in August 2025, tells you to type deepseek/deepseek-chat-v3-0324:free. On 2 October 2026 OpenRouter listed 17 models with a free version and none of them was a DeepSeek model. The paid version of that model is still listed, so the same name without :free works if you have credits.
Is Claude or ChatGPT a good model for Janitor AI?
They work, but few Janitor users choose them. Neither appears in the 20 models Janitor users run most on OpenRouter. Price is one clear reason: the same 200-turn session that costs about $2.50 on DeepSeek V4.1 Flash costs about $17 on Claude Sonnet 5.5 before caching. If you love how Claude writes and accept the cost, it is a fair choice, but it is not the common one.
Do I need janitor+ to use a better model on Janitor AI?
No. Connecting your own proxy or API key is free on Janitor’s side, on both the free plan and janitor+. You pay only the provider. janitor+ adds the Router, which lets you pick outside models inside Janitor without a proxy and pay per token from a wallet, and it gives Janitor’s own model five times more context.
Why does my Janitor AI proxy model show its thinking or reply slowly?
Many current models think before they answer, and on several of them thinking is on by default. Thinking adds waiting time, it is billed as output, and some hosts send the reasoning text into the chat. If your provider lets you turn thinking off, do that for fast scenes. If it does not, pick a model where thinking is optional.
The usage figures are from OpenRouter’s public app page for Janitor AI (“Top models used by Janitor AI this month”, last 30 days), read on 2 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 are from OpenRouter’s documentation on limits. JLLM, the proxy quickstart, the temperature advice, the token advice and the Router quotations are from Janitor’s help centre. The share of DeepSeek in the top 20, 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.
Product names and trademarks are the property of their respective owners. Arcanum is not affiliated with Janitor AI, OpenRouter, DeepSeek, NVIDIA or any model maker named here.