OpenRouter Free Models for AI Roleplay (2026): Limits, Privacy and Setup

Editor's note: limits and model list checked September 2026. Free models change often, so check openrouter.ai before relying on any of them.

Key takeaways

• Free model IDs end in :free (for example qwen/qwen3.8-27b:free).
• Limits: 20 requests/minute; 50/day under 10 credits purchased; 1,000/day with 10+ credits purchased. Going over returns a 429 error.
• Privacy: OpenRouter has separate settings for paid and free models about routing to providers that may train on your data. Check them.
• One roleplay reply = one request, and regenerations count too, so 50/day runs out fast.
• If you want to skip keys entirely, apps with built-in models are simpler.

What is OpenRouter?

OpenRouter is a router for AI models. Instead of signing up separately with each AI company, you create one OpenRouter account and one API key, and choose a model per request. It uses the common "OpenAI-compatible" format, so most chat apps that accept a custom API can use it.

In character-chat communities it's popular because:

• It's the most common way to use a proxy on Janitor AI (Janitor calls any custom API a "proxy").
• It offers free model variants, useful for testing.
• It works with SillyTavern and other self-hosted front ends.

OpenRouter free models: the limits in plain language

From OpenRouter's documentation at the time of writing:

Your account	Free-model requests per minute	Free-model requests per day
Bought less than 10 credits	20	50
Bought 10 or more credits	20	1,000

What that means for roleplay:

• Every reply is one request. So is every regenerate or "swipe".
• 50 a day is roughly one short scene. 1,000 a day is plenty for most people.
• To unlock 1,000/day you need to buy at least 10 credits once. That's a one-time top-up (roughly $10 at the time of writing, about ₹840), and it doesn't have to be spent on free models.
• Hitting the cap returns HTTP 429 (Too Many Requests). Wait for the reset, or switch to a paid model.

You can check your remaining allowance via OpenRouter's key endpoint (GET /api/v1/key), which reports your limits and free-model daily usage.

Which OpenRouter free models are available? (September 2026)

When we checked OpenRouter's public model list, there were 21 :free models. These general-purpose ones are the most suitable for conversation and storytelling:

Model ID	Context window	Notes
qwen/qwen3.8-27b:free	262K	Strong multilingual model; good Hindi potential
google/gemma-4-31b-it:free	262K	Google's open model; balanced and fluent
google/gemma-4-26b-a4b-it:free	262K	Lighter, faster Gemma variant
z-ai/glm-5.2:free	32K	Capable general chat; smaller context
nvidia/nemotron-3-super-120b-a12b:free	262K	Large model; can be slower
nvidia/nemotron-3-ultra-550b-a55b:free	1M	Very large; availability can fluctuate
thinkingmachines/inkling:free	~1M	General-purpose; long context

Several other free models are specialised (coding, finance, content-safety classification or vision) and aren't a good fit for chat.

Important caveats:

• The list changes. Models are added and removed, sometimes with little notice. If a model suddenly errors, check whether its :free variant still exists.
• Free capacity is shared. Popular free models can be slow or temporarily unavailable at peak times.
• Context isn't memory. A large context window helps, but the chat app decides how much history it sends.

Privacy: check this before you chat

OpenRouter's logging and privacy settings let you decide whether your requests can be routed to providers that may train on your data. There are separate settings for paid and free models. If you opt out, OpenRouter won't route to providers that train, which can reduce the free models available to you. Those settings control which providers you use; they don't change OpenRouter's own policies.

Practical rules:

• Don't paste personal information (phone, address, Aadhaar, PAN, bank details) into any AI chat.
• Assume free endpoints are less private than paid ones, unless your settings say otherwise.
• Never share your API key. Anyone with it can spend your credits. If it leaks, revoke it and create a new one.

How an OpenRouter key plugs into a chat app

The details differ by app, but almost every "custom API" or "proxy" screen asks for the same three things:

1. API URL / endpoint: OpenRouter's OpenAI-compatible base URL (shown in OpenRouter's quick-start docs).
2. API key: created in your OpenRouter account under Keys.
3. Model name: the exact ID, including :free if you want the free variant.

On Janitor AI: open the API settings in a chat, choose the custom/proxy option, and enter those three values. Janitor itself still stores your chats; OpenRouter only generates the replies. See our Janitor AI review for how JLLM and proxies compare.

On SillyTavern: choose Chat Completion, select OpenRouter as the source, paste your key and pick a model.

Always follow each app's own current instructions, since menus move.

Common OpenRouter errors and what they mean

Error	Meaning	Fix
401	Invalid or missing key	Re-copy the key with no spaces; create a new one if revoked
402	Not enough credits (paid models)	Add credits, or switch to a :free model
429	Rate limit hit (20/min or 50/1,000 per day)	Slow down, wait for the daily reset, or buy 10 credits to raise the daily cap
Model not found	Wrong ID or the free variant was removed	Check the model page and copy the exact ID
Empty or cut-off replies	Max tokens too low, or the provider timed out	Raise max response length; try another model

More Janitor-specific troubleshooting: Janitor AI not working? Fixes.

Tips for better roleplay on free models

• Keep bot definitions tight. Long personalities eat context and requests. Our character cards guide shows how to write a lean one.
• Regenerate less; steer more. A short (OOC: make Meera more sarcastic) costs one request and often beats three swipes. More in our AI roleplay chat guide.
• Set a sensible reply length (for example 250–400 tokens) so replies don't get cut off.
• Have a backup model. When your favourite free model is busy, switch instead of retrying.

The simpler route: no API keys at all

OpenRouter is great if you enjoy tinkering. If you'd rather just chat, especially in Hindi or Hinglish, an app with built-in models saves the hassle.

On Kirdaar AI (our app; full disclosure):

• You pick Fast, Balanced or Best quality instead of model IDs.
• There are no rate-limit errors to manage. You get free starter credits, then one-time packs from $5 (about ₹420) for 100 credits that never expire. See pricing.
• If you do like bringing your own account, you can add a Hugging Face token and your usage is billed to your Hugging Face account instead of Kirdaar credits. Free HF accounts get a small monthly inference credit ($0.10 at the time of writing, per Hugging Face's pricing docs); after that, HF bills at provider rates.
• Replies in English, Hindi or Hinglish, switchable per chat, with Quick chat, Roleplay, Storytelling and Deep talk modes.

The honest trade-off: Kirdaar has a smaller character library and fewer model choices than OpenRouter + SillyTavern. Compare in Kirdaar AI vs Janitor AI.

You: Kahani shuru karo: ek purani haveli, barish, aur ek band darwaza.

Kahaanikaar (Storytelling, Hinglish): Barish ki boondein haveli ki tooti chhat se tapak rahi hain. Lantern ki roshni mein tumhe ek lohe ka darwaza dikhta hai, jispe kisi ne chaak se likha hai: "Mat kholna." Tum kya karoge? 1) Darwaza khologe 2) Chaak ka likha padhoge 3) Wapas mudoge.

How to get started on Kirdaar AI

1. Go to kirdaarai.com and sign up for free starter credits.
2. Choose English, Hindi or Hinglish.
3. Pick a character or create your own.
4. Choose a mode and a quality setting. That's it: no URLs, keys or model IDs.

Bottom line

OpenRouter's free models are a genuinely useful playground: one key, many models, zero cost, as long as you respect the limits and check your privacy settings. If you'd rather skip the setup and chat in Hinglish straight away, try Kirdaar AI with free starter credits.

18+. Characters are AI. Everything they say is made up.