What’s Changed: Chutes raised its Janitor AI proxy prices across several models with no announcement, with DeepSeek 3.2 jumping from $0.24/$0.42 to $1.00/$1.00 per million tokens. The proxy that used to be the cheapest pick is now mid-pack. This guide covers what changed, the cheapest model and proxy combinations right now, and how to switch without losing your chats.
If you run roleplay on Janitor AI through a proxy, the Janitor AI Chutes proxy price you signed up for may not be the price you are paying today.
Chutes quietly pushed through a round of increases across several popular models, and the first sign most people got was a credit balance draining faster than usual.
I have watched Chutes go from the default budget recommendation to a provider people are leaving in a single week. The frustrating part is that none of this came with a heads up, so users only noticed when their wallet numbers stopped adding up. That is a bad way to find out your cost per message doubled.
Here is the good news. The model you want for roleplay is cheaper than ever from other providers, one strong option is genuinely free, and moving off Chutes takes about five minutes once you know which fields to change. You will finish this knowing exactly what to switch to and how to keep from overpaying again.
What is BYOK: Bring Your Own Key means you supply your own API key from a model provider, and Janitor AI sends your chats through that key instead of its own backend.
What Changed With Janitor AI Chutes Proxy Pricing
The Janitor AI Chutes proxy price increase raised DeepSeek 3.2 from $0.24/$0.42 to $1.00/$1.00 per million tokens, with GLM models up 25 to 50 percent, and it shipped with no announcement.
That is roughly a 2x to 4x jump on the model most roleplay users had standardized on.
From what I have seen, the GLM family moved too. GLM-5 went from about $0.60/$1.92 to $0.95/$2.55 per million input and output tokens, and GLM-5.1 climbed from $0.98/$3.08 to $1.20/$4.00.
None of those numbers are catastrophic on their own, but together they push Chutes out of the cheapest tier it used to own.
Two other changes stung longtime users. Chutes retired older models people relied on, including the trusty R1-0528, and it dropped the one time payment option that budget users liked.
The context matters here because DeepSeek’s whole appeal was undercutting everyone, as Statista’s DeepSeek-R1 price chart shows.
If your chats started throwing connection problems around the same time, that is a separate issue worth ruling out, and our walkthrough on Chutes fetch errors covers that path.
Why the Price Hike Hits Roleplay Users Hardest
Roleplay burns more tokens than almost any other AI use case, so a per token increase compounds fast across a long story.
A single immersive session can run tens of thousands of tokens, and every swipe resends the full context.
The way I see it, the cruelest detail is what the increase does to caching math. DeepSeek charges a tiny fraction for repeated content, which means your character card and system prompt normally cost almost nothing on each new message. When the base rate quadruples, that hidden discount shrinks in raw terms too, so the model you tuned around cheap caching is no longer the bargain you planned for.
Long sessions also trigger a trap most people never read about. Some flagship models charge a long context penalty once a chat passes a threshold, billing input at 2x to 4x the normal rate past 200,000 tokens. Pair that with a proxy hike and a marathon roleplay can cost several times more per reply at message 300 than it did at message 10.
What is context caching: Caching stores the repeated front of your prompt, like your character definition, so the provider charges a steep discount for it instead of full price on every message.
The Cheapest Janitor AI Proxy and Model Options Now
The cheapest reliable setup right now is DeepSeek V4 Flash through the direct DeepSeek API, with Z.ai’s GLM-4.7-Flash as a genuinely free backup.
Going direct skips the proxy margin and the extra network hop, which means lower cost and faster replies.
I would start with the table below before committing to any one provider. Prices move, so treat these as the current shape of the market rather than permanent figures.
| Provider | Best roleplay model | Cost per 1M (in / out) | Best for |
|---|---|---|---|
| DeepSeek Direct | V4 Flash | $0.14 / $0.28 | Lowest cost, fastest, 1M context |
| Z.ai Direct | GLM-4.7-Flash | Free | Zero cost backup, light sessions |
| NanoGPT | DeepSeek, GLM variants | $8/mo for ~60,000 queries | Flat monthly, model switching |
| OpenRouter | DeepSeek V4, 59 free models | List price, free tier available | Flexibility, trying many models |
| Featherless | DeepSeek, Kimi, GLM | $10 to $25/mo flat, unlimited | Heavy users who hate metering |
| Chutes (new price) | DeepSeek 3.2 | $1.00 / $1.00 | No longer the budget pick |
A few details that do not fit in a table but change the decision. NanoGPT advertises no markup pricing, yet it adds a 5 percent fee if you manually pin a specific provider or use a BYOK request, so the default routing is the cheap path.
Featherless looks unbeatable at a flat rate, but the $10 Basic tier caps you at two concurrent connections and a 16,000 token context window, which is tight for long roleplay.
OpenRouter has a quiet pro tip worth knowing. Once you buy $10 in credits a single time, your daily limit jumps from 50 to 1,000 requests, and you keep that higher limit even if your balance later dips below $10.
For a deeper look at running DeepSeek without hitting walls, our guide on unlimited messages on DeepSeek goes further.
How to Check What You Are Being Charged
The fastest way to catch a stealth price hike is to read your provider’s usage dashboard and confirm your three connection fields all belong to the same provider.
Most billing surprises trace back to a setting you can verify in under a minute.
In my experience, the errors people panic about are usually money problems wearing a disguise. The table below maps the symptom to the real cause so you stop guessing.
| Symptom | Likely cause | Fix |
|---|---|---|
| Credits draining faster than last week | Provider raised per token price | Check the dashboard rate, compare against the table above |
| “Network Error” mid chat | Zero balance returning a 402 payment error | Top up credits or switch to a free model |
| “Network Error” right after setup | Key and URL belong to different providers | Match all three fields to one provider |
| 404 Model Not Found | Model name typo or wrong namespace | Copy the exact model slug from provider docs |
| Replies suddenly more expensive late in a chat | Long context penalty past the threshold | Trim the chat or summarize older turns |
What is the three field rule: Janitor AI needs a Proxy URL, an API Key, and a Model Name that all come from the same provider. Mix a DeepSeek key with an OpenRouter URL and you get a generic Network Error.
Here is the sequence I walk through whenever my numbers look off.
- Open your provider’s usage or billing page and read the current per million token rate, not the rate you remember.
- Confirm your balance is above zero, since an empty balance shows up in Janitor AI as a vague Network Error.
- Check that your Proxy URL, API Key, and Model Name all come from one provider.
- Copy the exact current model slug from the provider’s docs to avoid a 404.
- If the rate climbed, move your key to a cheaper provider from the table before your next session.
One forward looking note while you are in there. DeepSeek is retiring its legacy model aliases deepseek-chat and deepseek-reasoner on July 24, 2026, so update your Model Name field to a current name like deepseek-v4-flash to avoid a sudden outage.
If your proxy throws connection errors rather than billing ones, our piece on Janitor AI proxy connection failures is the better starting point.
How to Switch Proxies Without Losing Your Chats
Switching providers in Janitor AI only changes the model behind your chats, not the chats themselves, so your history and characters stay intact.
The proxy is the engine, not the save file.
What I would do is line up the new key first, then swap. That way you are never stuck mid scene with a dead connection.
- Create an account with your chosen provider and generate an API key.
- Add a small amount of credit, or pick a free model if you want to test first.
- In Janitor AI, open the proxy settings and paste the new Proxy URL, API Key, and Model Name together.
- Send one test message in a low stakes chat to confirm replies come through.
- Return to your main roleplay, where the character and history are exactly as you left them.
Before: A 50 message session on DeepSeek 3.2 at the new Chutes rate of $1.00/$1.00 per million tokens, with heavy context resending, can run a few dollars across the session.
After: The same 50 messages on DeepSeek V4 Flash direct at $0.14/$0.28, with caching cutting the repeated character card to a fraction of a cent per turn, often lands under a quarter of that cost.
That gap is why so many users are moving. The model quality is comparable, and the bill is a fraction of what the hiked proxy now charges.
When Skipping Proxies Altogether Makes Sense
If juggling API keys, credits, model slugs, and surprise price hikes is the part you hate, a companion app that bundles the model removes all of it.
You trade some control for never touching a billing dashboard again.
I will be honest about the tradeoff. Proxies give you raw model choice and the lowest possible per token cost, and for tinkerers that is the whole appeal.
For everyone else, the maintenance tax is real, and a self contained app means no proxy URL, no key rotation, and no quiet rate changes draining your wallet.
If that sounds like you, Candy AI runs its own model with memory built in, so there is no proxy to configure and nothing to top up mid chat.
For users who want strong character recall without the setup, Nectar AI is the alternative I would point to. Both skip the entire proxy headache that started this whole migration.
If you would rather stay in the open model world, our roundup of Janitor AI alternatives covers other routes, and a local model like Gemma 4 12B now runs in 16GB of RAM for a free, private option if you have the hardware.
Frequently Asked Questions
Is Chutes still the cheapest Janitor AI proxy?
No. After the recent increase, Chutes DeepSeek 3.2 sits at $1.00/$1.00 per million tokens, while DeepSeek V4 Flash direct runs $0.14/$0.28. Going direct or using NanoGPT is now cheaper for most roleplay users.
Will blocking Chutes as a provider make my model cheap again?
Only if you route to a cheaper provider for the same model. Blocking Chutes on an aggregator pushes traffic to the next available provider, so confirm where your requests land and what that provider charges.
What is the cheapest model for Janitor AI roleplay right now?
DeepSeek V4 Flash is the budget standard at $0.14/$0.28 per million tokens, with a 1 million token context window. Z.ai’s GLM-4.7-Flash is free and works well for lighter sessions.
Why did I get a Network Error after the price change?
A Network Error in Janitor AI often means a billing failure, not a connection problem. An empty balance returns a payment error, and a mismatched key and URL returns an authorization error, both shown as a generic Network Error.
Do I lose my chats if I switch proxies?
No. The proxy only powers the responses, so your chat history and characters stay exactly where they are. You paste new connection details and continue the same conversation.
Quick Takeaways
- Chutes raised DeepSeek 3.2 from $0.24/$0.42 to $1.00/$1.00 per million tokens with no announcement, so it is no longer the budget pick.
- DeepSeek V4 Flash direct at $0.14/$0.28 is the cheapest reliable setup, and Z.ai’s GLM-4.7-Flash is a free backup.
- A Janitor AI Network Error is usually a billing or field mismatch issue, not a connection fault.
- Switching proxies keeps your chats intact, since the proxy only powers responses.
- If you are done managing keys and credits, a bundled companion app like Candy AI removes the proxy layer entirely.
