The one fact that breaks every old tutorial
Setting up DeepSeek on Janitor AI changed on July 24, 2026: DeepSeek officially discontinued deepseek-chat and deepseek-reasoner, the two model names that every tutorial, and JanitorAI’s own help-center guide from May 2025, still tells you to type. Point a proxy at them now and the reply never comes; Janitor shows its generic “network error” and you start debugging the wrong thing.
The current model names are deepseek-v4-flash (public beta since July 31, 2026) and deepseek-v4-pro. Everything below uses them, through the two routes that actually work today, with what each costs. If you have never configured a proxy on Janitor at all, do my Janitor AI proxy setup first; this page assumes you know where the API Settings screen lives.
Route 1: DeepSeek direct (official, $2 minimum)
The straight line. You get a key from DeepSeek, fund a small balance, and point Janitor at DeepSeek’s endpoint.
- Create an account at platform.deepseek.com and generate an API key.
- Add balance. The minimum top-up is $2; $5 buys real headroom.
- In your Janitor chat: menu, API Settings, Proxy, Custom.
- API URL:
https://api.deepseek.com/v1/chat/completions - Model name:
deepseek-v4-flash(exactly; the old names are dead) - Paste the key with no trailing spaces, save, and hard-refresh the page.
- Send a real message. The check button is unreliable; the message is the test.
What it costs, from the official pricing page captured August 2, 2026: v4-flash runs $0.14 per million input tokens and $0.28 per million output; v4-pro runs $0.435 and $0.87. DeepSeek has also announced a future peak-hours surcharge of up to 2x, so treat evening pricing as potentially double. Even doubled, a chat-heavy month rarely dents a $5 balance, which is why the community’s standing verdict on DeepSeek is “just pay.”
Route 2: OpenRouter (cheaper per token, one key for everything)
Counterintuitive but true at capture: OpenRouter serves DeepSeek v4-flash for less than DeepSeek does, $0.09 in and $0.18 out per million on deepseek/deepseek-v4-flash-0731. OpenRouter also gives you one key that reaches hundreds of other models, which matters the next time a provider retires something overnight.
- Key from openrouter.ai (starts
sk-or-v1-), fund any small amount. - API URL:
https://openrouter.ai/api/v1/chat/completions - Model name:
deepseek/deepseek-v4-flash-0731, copied from the live model page, never typed from a guide, including this one. - Save, hard-refresh, send a message.
The free DeepSeek is gone, and the guides will not tell you
The route half the internet still teaches, deepseek/deepseek-chat-v3-0324:free on OpenRouter, no longer exists. I pulled OpenRouter’s live model index on August 2, 2026: 337 models, 14 with the :free suffix, and not one of them is DeepSeek. JanitorAI’s own three help articles all still teach the free-DeepSeek route by name. And the “new accounts get 5 million free tokens” line you will meet on some blogs is contradicted by DeepSeek’s own pricing and updates pages, which offer no free tier and no trial credit.
If free is the requirement, the honest answer moved: the surviving free lanes are non-DeepSeek models, and my free uncensored models ranking measures which of them are worth the rate limits. If DeepSeek’s writing style specifically is what you want, it costs money now; the good news is how little.
When it still says “network error”
Janitor flattens upstream failures into one generic message, so the same words mean three different problems. In order of 2026 likelihood:
| Likely code | What actually happened | Fix |
|---|---|---|
| Dead model | The name in your settings was retired July 24 | Replace with deepseek-v4-flash or the OpenRouter string above |
| 401 | Wrong key, a stray space, or a key from the other provider | Re-paste; check the key prefix matches the URL |
| 402 | DeepSeek balance at zero | Top up; free tokens do not exist |
| 429 | Rate or quota pressure | Wait it out, or move to the other route |
The full error-code table for every provider, not just DeepSeek, lives on my proxy guide; if the whole site is misbehaving rather than one chat, check whether Janitor is actually down before touching your settings.
If the setup dance is the problem
DeepSeek-on-Janitor is the best free-platform-plus-cheap-model combination going, and it now involves a payment method, model names that changed twice in a summer, and an error message that refuses to explain itself. If what you wanted was uncensored chat that just works when you open it, that is a different product category, and it is the one I actually test: Candy AI was the strongest texting of my test cycle on a flat plan, and my JanitorAI alternatives page ranks the platforms by what pushed you off Janitor in the first place.
Sources
- DeepSeek API updates, model discontinuation notice, captured Aug 2, 2026
- DeepSeek API pricing, v4 prices, captured Aug 2, 2026
- OpenRouter models API, live free-model census and v4-flash listing, queried Aug 2, 2026
- JanitorAI help center proxy guides, the official instructions this page updates, accessed Aug 2, 2026