DeepSeek on Janitor AI: the proxy fields, every DeepSeek build, and the price per message

Open a chat whose creator allows proxies. Click the gear or menu icon at the top right, choose API Settings, pick Proxy under Provider, then + Add Configuration and fill in:

• Model Name: a DeepSeek id from the table, pasted exactly. Janitor sends it as the model field, so a display name such as "DeepSeek V4.1" or a stray space gets a 404.
• Proxy URL: https://api.relayswitch.io/v1/chat/completions, the full path. A URL that stops at /v1 fails.
• API Key: the key, with no space before or after it.
• Custom Prompt: leave it blank. The character card and your persona already go with every message, and a prompt here is billed as input on each turn.

Click Add, then Save Settings, refresh the page, and check the configuration reads Active. The first message is the real test; Janitor's own quickstart says to skip the Check Key/Model button. The JanitorAI proxy guide walks every field and every error code; this page covers the DeepSeek part.

Every DeepSeek id on the key, with the kind of build, its context window, and the live price of one 2K-in / 300-out turn and 100 of them. The builds tagged uncensored or abliterated come first: they are DeepSeek models with the refusal behaviour tuned out or removed from the weights, so a card plays as written without a jailbreak prompt riding along as paid input. The uncensored models page shows each one's daily check. A $10 top-up covers about 4,531 such turns on DeepSeek V4.1 Flash Abliterated.

A build marked "thinks first" reasons before it answers and spends part of the reply budget on that. If its replies come back short or cut off, raise max tokens in Janitor's generation settings to 1,000 or more, or pick the same build without the thinking step. The best models for Janitor AI list ranks every chat model on the key by use in the last seven days, DeepSeek among them.

DeepSeek's own API serves the stock models only, and Janitor's proxy form takes one URL and one key. Through RelaySwitch the same two fields reach the stock builds, the uncensored and abliterated builds, and every other model on the Models page, so switching from DeepSeek V4.1 Flash to V4 Pro, or to a Claude or Gemini model, is one change of Model Name. A failed request is handed to the next host for that model at the same price, and every reply carries a signed receipt you can check on the verify page.

To confirm the key sees the DeepSeek ids before you paste one into Janitor:

Menu path, field names, the full-URL rule and the save-and-refresh flow are from Janitor's help center quickstart, as read for the JanitorAI proxy guide on 2026-11-02. Model ids, tags, context windows and prices come from the RelaySwitch catalog at render. The same key runs DeepSeek in SillyTavern, Chub AI and RisuAI.