OpenRouter is the simplest way to try many AI models without signing up to many providers. One account and one API key give you more than 600 models from over 80 providers, including OpenAI, Anthropic, Google, DeepSeek, Qwen, Mistral and Nvidia, through a single OpenAI-compatible endpoint. It also has a set of free models that are great for testing and hobby projects. The catch is the details: a 5.5% fee on credits, tight free limits that change based on whether you have ever bought $10 of credits, and free models that come and go every few weeks. This guide covers all of it as of November 2026.
- What it is: an API router. Same request format, any model: change the
modelstring and you switch providers. - Cost: provider list prices with no token markup, plus a 5.5% fee on credit purchases (Standard plan). Business plan fee is 8%.
- Free tier: 25+ free models, 20 requests per minute, 50 per day, or 1,000 per day once you have bought $10 of credits (one time).
- BYOK: use your own provider keys; the first $25,000 of inference per month has no fee, then 5%.
- Free models now include several Nvidia Nemotron models, Cohere North Mini Code and LiquidAI LFM2.5. DeepSeek R1, Qwen3 Coder and Llama are no longer free.
What Is OpenRouter?
OpenRouter sits between your app and the model providers. You send an OpenAI-style chat request to https://openrouter.ai/api/v1, pick a model by its ID in provider/model form (for example a Claude, Gemini or DeepSeek model), and OpenRouter forwards it to a provider that serves that model, falling back to another provider if the first is down. You get one bill, one dashboard and one set of logs.
Why people use it:
- Compare models quickly without creating accounts at OpenAI, Anthropic, Google and a dozen open-model hosts.
- Uptime: automatic fallback across providers for the same open model.
- Free models for learning and prototypes.
- Works with existing tools: any app that accepts a custom OpenAI base URL, including Cline, Continue, n8n and many chat UIs.
OpenRouter Plans and Fees
| Plan | Fee | Models | Highlights |
|---|---|---|---|
| Free | None | 25+ free models from 4 providers | 50 requests/day, activity logs, no BYOK |
| Standard (pay as you go) | 5.5% on credits | 600+ models, 80+ providers | Provider rate limits pass through, prompt caching, budgets and spend controls |
| Business | 8% | Same | Up to 1,000 workspaces, EU/US in-region routing, advanced notifications |
| Enterprise | Custom, discounts | Same | SSO/SAML, SLAs, invoicing, private endpoints add-on |
Token prices themselves match each provider’s list price; the fee is charged when you buy credits. Payment works by card or crypto. For the underlying model prices, see my AI API pricing comparison.
Free Tier Limits Explained
| Situation | Requests per minute | Requests per day (free models) |
|---|---|---|
| Never bought credits | 20 | 50 |
| Bought at least $10 in credits (one time, lifetime) | 20 | 1,000 |
| Balance below zero | Free models stop with a 402 error until you top up | |
The $10 is not a subscription; it is a one-time threshold. Buying more does not raise the per-minute limit. Upstream providers can still rate-limit free models at busy times. Check your status with GET https://openrouter.ai/api/v1/key: the is_free_tier field turns false once you have bought credits.
My advice: if you are building anything real, buy the $10 once. It raises your daily free allowance twentyfold and the credits still work on paid models.
Free Models on OpenRouter (November 2026)
Free models change often. This list comes from a tracker that tested them on November 4, 2026:
| Model | Good for | Status in test |
|---|---|---|
| NVIDIA Nemotron 3 Ultra | General chat, 1M context | Answered |
| NVIDIA Nemotron 3.5 Super VL | Vision plus text, 1M context | Answered |
| NVIDIA Nemotron 3.5 Lightning | Fast replies | Answered |
| NVIDIA Nemotron 3 Nano Omni | Small multimodal | Answered |
| Cohere North Mini Code | Coding | Answered |
| XingChen Xing4.0 29B A4B | General chat | Answered |
| LiquidAI LFM2.5-2.6B | Tiny, fast tasks | Answered |
| Dots3-Note Preview, Recursive Flint | General chat | Answered |
| Google Gemma 4 26B A4B and 31B | General chat | 429 from Google AI Studio during test |
Easier option: use the model ID openrouter/free, which picks an available free model automatically. Older favourites such as DeepSeek R1, Qwen3 Coder and Llama models are no longer free. If you want a model that is always free and never rate-limited, run it locally: my best Ollama models guide and Gemma 4 local guide show how.
How to Start: Key, Code and Tools
- Sign up at openrouter.ai (no card needed for free models).
- Create a key under Keys. Set a credit limit on the key if it will live in an app.
- Call it with any OpenAI SDK by changing the base URL:
pip install openai
from openai import OpenAI
client = OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="sk-or-...",
)
resp = client.chat.completions.create(
model="openrouter/free", # or e.g. "anthropic/claude-sonnet-5.5"
messages=[{"role": "user", "content": "Explain RAG in 3 sentences"}],
max_tokens=1000,
)
print(resp.choices[0].message.content)Copy exact model IDs from each model’s page on OpenRouter, since names change with versions. In tools like Cline, n8n or Open WebUI, choose “OpenAI compatible”, set the base URL above and paste your key. My n8n AI agent guide shows the same pattern with a local model.
How to Pick a Free Model That Actually Answers
- Open the Models page on OpenRouter and narrow the list to free models (their IDs end in
:free). - Check the context length and whether the model accepts images, so it fits your task before you write any code.
- Send a short test prompt from the model’s playground. If it returns a 429 from the upstream provider, try the next model rather than retrying the same one.
- Keep two or three working free model IDs in your config and fall back between them, or simply use
openrouter/freeand let OpenRouter choose. - Re-check your list every few weeks, because free models are added and removed often.
How to Set Spend Limits Before You Ship
Once you buy credits, a leaked key or a runaway loop can spend real money. OpenRouter’s spend controls let you cap that:
- Create a separate key for each app or environment, so one leak does not expose everything.
- Create a guardrail with a dollar limit and a reset interval (daily, weekly or monthly), then assign it to the key.
- If you work in a team workspace, also set a workspace budget, which caps total spend across all members and keys.
- Decide whether BYOK spend should count. The “include BYOK spend” toggle is off by default, so only OpenRouter credit spend counts toward the limit.
- Remember that raising a limit does not reset the counter; it rolls over only at the next interval.
Model Suffixes and Routing
:freecalls the free variant of a model, for examplenvidia/nemotron-3-super-120b-a12b:free. Not every model has one.:nitroprefers the fastest providers.:floorprefers the cheapest providers.:exactoprefers providers with the best tool-calling quality.:thinkingand:onlineare deprecated, and:extendedno longer works.
BYOK: Bring Your Own Key
On Standard and Business you can add your own provider keys (for example your Anthropic or Google key). Requests then bill to your provider account, and OpenRouter charges no fee on the first $25,000 of list-price inference per month, then 5%. BYOK is also a good fix for free-model 429 errors, because requests draw on your own provider quota instead of the shared pool. BYOK is not available on the Free plan.
Common Errors and Fixes
| Error | Meaning | Fix |
|---|---|---|
| 429 “rate-limited upstream” | The provider behind a free model is busy; your daily quota is not used | Switch model or provider, use openrouter/free, or add BYOK |
| 429 on your own limit | You hit 20/min or your daily cap | Slow down, or buy $10 once for 1,000/day |
| 402 | Negative balance | Top up credits |
| 403 “agentic harnesses” | Model only works inside agent tools | Use a different model for plain API calls |
HTTP 200 but no choices |
Upstream error inside the response body | Check for an error field before reading choices |
| Empty reply | Reasoning used up max_tokens |
Raise max_tokens to around 1,000 or more |
Similar errors on Anthropic’s own API are covered in my Claude 429 and 529 fix guide.
Is OpenRouter Worth It vs Going Direct?
Use OpenRouter when you are testing several models, need automatic fallback for open models, or want one bill. The 5.5% fee is small next to the time saved. Go direct when you have settled on one provider and spend a lot, because you avoid the fee and get provider-only features first. A common middle ground is OpenRouter for experiments and BYOK for production. If cost is the main concern, my DeepSeek API pricing guide shows one of the cheapest strong options.
Frequently Asked Questions
Is OpenRouter free?
Partly. It has 25+ free models with 20 requests per minute and 50 per day, or 1,000 per day after a one-time $10 credit purchase. Paid models cost provider list price plus a 5.5% credit fee.
Does OpenRouter add a markup to token prices?
No token markup; the fee is charged on credit purchases (5.5% on Standard, 8% on Business).
Which free model is best on OpenRouter?
It changes weekly. In early November 2026 the Nvidia Nemotron 3 Ultra and 3.5 Super VL models were reliable choices, with Cohere North Mini Code for coding. Or use openrouter/free to auto-pick.
Is OpenRouter safe for private data?
Requests go to third-party providers. Standard and above let you route by data policy, and Business adds EU/US in-region routing. For sensitive data, check each provider’s policy or run models locally.
How do I stop OpenRouter from spending too much?
Use a separate key per app and attach a guardrail with a dollar limit that resets daily, weekly or monthly. Teams can add a workspace budget on top. Keep your balance small while testing.
Why do I get 429 errors on a free model when I am under my limit?
The provider behind that free model is busy, so the request was rate-limited upstream. Your own daily count is not the cause. Switch to another free model, use openrouter/free, or add your own provider key with BYOK.
