Step 1: Create an account and an API key
- Create a free TokenReply account. Free models work right away, no payment needed.
- Create an API key on the API keys page (walkthrough) and copy the value that starts with
sk-.
Step 2: Add TokenReply as a proxy in JanitorAI
- Open any chat on JanitorAI, select the gear icon at the top right, then API Settings.
- Choose the proxy option and select Add Configuration.
- Config Name: any name, for example
TokenReply. - Model Name: an exact model ID from the list below, for example
deepseek-v4-flash. Type it without spaces. - Proxy URL:
https://api.tokenreply.com/v1/chat/completions. Use the full endpoint, not just the base URL. - API Key: your
sk-key, with no spaces before or after it. - Custom Prompt: optional. Leave it empty unless you already use one.
- Select Add, then Save Settings, then refresh the page. JanitorAI only loads the new configuration after a refresh.
- Send a short test message in a chat with a character that allows proxies.
Step 3: Pick a model for roleplay
These model IDs answered in character with JanitorAI's usual settings when we tested them. Prices and tiers below are read live from the catalog. Copy the ID exactly into Model Name; make one configuration per model to switch quickly.
Read from the live catalog when this page opens.
Access: Free models work on every account at no charge. Models of the week (Weekly Featured) allow 40 calls a day on Default and Trial accounts and have no daily cap on Plus. Plus models need a Plus subscription; calling them without one returns 403.
Free to try
Free models cost nothing per message and work on every account, subject to availability and rate limits.
| Model ID | Access | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|---|
nvidia/nemotron-3-ultra-550b-a55b | Free | Free | Free | Free |
grok-4.20-fast | Free | Free | Free | Free |
nemotron-3.5-super-vl-free | Free | Free | Free | Free |
gemma-4-31b-it | Free | Free | Free | Free |
Low cost
Well under a cent for a typical long-chat message, and cheaper still when the chat history is cached.
| Model ID | Access | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|---|
deepseek-v4-flash | Models of the week | $0.22 | $0.007 | $0.66 |
mimo-v2.5-pro | Models of the week | $0.217 | $0.0018 | $0.435 |
minimax-m2.5 | Models of the week | $0.09 | $0.009 | $0.36 |
glm-5.3-flash | Plus | $0.075 | $0.015 | $0.25 |
gemini-2.5-flash | Models of the week | $0.15 | $0.015 | $1.25 |
Balanced
Stronger writing and memory for long stories at a moderate price per message.
| Model ID | Access | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|---|
kimi-k2.7 | Models of the week | $0.475 | $0.095 | $2 |
glm-5.2 | Models of the week | $0.7 | $0.13 | $2.2 |
glm-5.3 | Plus | $0.7 | $0.13 | $2.2 |
deepseek-v4-pro | Plus | $0.66 | $0.022 | $1.98 |
mimo-v2.6-pro | Plus | $0.217 | $0.0018 | $0.435 |
Premium
The most capable models for complex characters and long arcs. Plus accounts only.
| Model ID | Access | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|---|
claude-sonnet-5 | Plus | $0.6 | $0.06 | $3 |
claude-opus-4-8 | Plus | $1.5 | $0.15 | $7.5 |
claude-opus-5 | Plus | $1.5 | $0.15 | $7.5 |
Recommended JanitorAI settings
Generation settings that work well with TokenReply
| Setting | Recommended value | Why |
|---|---|---|
| Text streaming | On | Text appears as it is written, and long replies are not cut off by the timeout described below. |
| Temperature | 0.8 to 1.0 to start | Higher values are accepted, and TokenReply adjusts sampling values a model's provider would reject, so JanitorAI's defaults do not cause errors. |
| Max new tokens | 600 to 1000 | Long enough for a full roleplay reply. Avoid 0 or extreme values: a few models reject them with a 400 error. |
| Context size | 16k to 32k tokens for most chats | Every message resends the chat up to this size, so a larger context costs more per message. Cache makes the repeated part cheaper. |
Timeouts: keep streaming on
A reply that is sent all at once (streaming off) must finish within about 2 minutes, or the connection ends with a timeout error and nothing is shown. With text streaming on, long replies keep arriving until they are complete. If you use a VPN and replies stop after about two minutes even with streaming on, the VPN may be closing long connections; try without it or with a lower Max new tokens.
What does one message cost?
Read from the live catalog when this page opens.
Example: one message in a long chat sends 16,000 tokens (character, memory and history) and gets 800 tokens back. With cache, 12,000 of the 16,000 input tokens are the unchanged part of the chat and are billed at the cached input price. Free models cost nothing.
| Model ID | Per message | Per message with cache | 100 messages with cache |
|---|---|---|---|
deepseek-v4-flash | $0.00405 | $0.00149 | $0.149 |
mimo-v2.5-pro | $0.00383 | $0.00124 | $0.124 |
glm-5.3-flash | $0.0014 | $0.00068 | $0.068 |
kimi-k2.7 | $0.0092 | $0.00464 | $0.464 |
glm-5.2 | $0.013 | $0.00612 | $0.612 |
deepseek-v4-pro | $0.0121 | $0.00449 | $0.449 |
claude-sonnet-5 | $0.012 | $0.00552 | $0.552 |
claude-opus-4-8 | $0.03 | $0.0138 | $1.38 |
Prices use the cheapest available route, as on the models page. Actual cost depends on your context size, reply length and how much of the chat is cached.
Free models cost nothing per message. Weekly Featured models (labeled Models of the week) allow 40 calls a day on Default and Trial accounts, plus 10 more on days you check in. Plus ($5 a month, with $5 of usage included) unlocks Plus models and removes that daily cap. Details are in plans and tiers.
Troubleshooting
| Symptom | What it usually means | What to do |
|---|---|---|
A network error occurred or Network Error | The Proxy URL is not the full endpoint, or the page was not refreshed after saving. | Set Proxy URL to exactly https://api.tokenreply.com/v1/chat/completions, select Save Settings and refresh the page. |
401 Invalid token | The key is incomplete, revoked, or pasted with spaces. | Copy the key again from the API keys page and paste it into API Key without spaces. |
401 Token not provided | The API Key field is empty. | Paste your sk- key, save and refresh. |
503 No upstream channel is available for this model | The model ID is misspelled, has a trailing space, or is not in the catalog. | Copy the exact ID from the list on this page or the models page into Model Name. |
400 The request failed with status code 400 | Usually Max new tokens is 0 or very large, or the chat is longer than the model's context window. | Set Max new tokens to 600 to 1000 and lower the context size, then send the message again. |
403 or not permitted for a model | Your account tier does not include that model. | Pick a Free or Weekly Featured model, or upgrade to Plus. The tier of every model is shown on the models page. |
| 429 Too Many Requests | A rate limit or the daily Weekly Featured allowance was reached. | Wait a moment before retrying, or switch to another model. Plus has no daily Weekly Featured cap. |
502 or Service temporarily unavailable | The model is busy right now. | Retry the message. If it keeps failing, switch to another model from the list above and check status. |
| The reply stops mid-sentence | The reply reached Max new tokens. | Raise Max new tokens, or ask the character to continue. |
| Nothing appears after about 2 minutes | Streaming is off and the reply took longer than the timeout. | Turn on text streaming, or lower Max new tokens. |
| The old model is still used after saving | JanitorAI has not reloaded the configuration. | Refresh the page, open API Settings and check which configuration is active. |
Frequently asked questions
What is the JanitorAI proxy URL for TokenReply?
https://api.tokenreply.com/v1/chat/completions. JanitorAI needs the full chat completions endpoint. Apps such as SillyTavern ask for the base URL https://api.tokenreply.com/v1 instead.
Which model should I use on JanitorAI?
Start with a free model such as grok-4.20-fast or nvidia/nemotron-3-ultra-550b-a55b. For better writing at a low price, deepseek-v4-flash and mimo-v2.5-pro are good choices; claude-sonnet-5 is the premium pick on Plus.
Is there a free JanitorAI API?
Yes. Free models cost nothing per message on every TokenReply account, within rate limits. The current list is in the free AI API guide.
Why are long JanitorAI chats expensive, and how does cache help?
JanitorAI resends the character and the chat history with every message, so a long chat sends thousands of tokens each time. On models with a cached input price, the repeated part is billed at that lower rate.
How do I pay for paid models?
Add balance on the top-up page (how to top up) and pay as you go, or subscribe to Plus for $5 a month, which includes $5 of usage and unlocks Plus models. Free models never need a balance.
Do I need to refresh JanitorAI after changing settings?
Yes. Save the configuration, then refresh the page before sending a message.