OpenRouter is the easiest way to try free LLMs: one account, one key, and dozens of models priced at $0. But "free" on OpenRouter comes with a daily request cap, a per-minute cap, and a few settings that decide which free models you can call at all. Here is how it works in September 2026.
The limits in one table
| Account state | Requests per day on :free models | Requests per minute |
|---|---|---|
| New account, no purchases | 50 | 20 |
| At least $10 of credits bought (one time) | 1,000 | 20 |
The limits apply to all :free variants together, not per model. The $10 does not get spent on free models — it only unlocks the higher daily cap, and you can use it later for paid models. That makes the top-up the cheapest upgrade in the free-LLM world if you use OpenRouter daily. Source: OpenRouter limits documentation.
How :free models work
Many models on OpenRouter have two variants: a paid one (z-ai/glm-5.2) and a free one (z-ai/glm-5.2:free). The free variant is usually served by a provider that donates capacity or runs a promotion. Consequences:
- Capacity is shared. Popular free models return HTTP 429 at peak times even if you are under your daily cap.
- Terms differ per model. Some free endpoints may log or train on prompts. OpenRouter's privacy settings control whether such endpoints are allowed for your account; if a free model returns an error about data policy, that setting is the reason.
- Free variants come and go. Promotional and "stealth" models appear for a few weeks. Our new free models feed tracks them hourly.
Which free models are on OpenRouter now
On September 24, 2026 our catalog lists 24 free OpenRouter endpoints, including Nemotron 3 Ultra, Nemotron 3 Super, GLM 5.2, Qwen3.8 27B, Gemma 4 31B, Inkling from Thinking Machines Lab, and Poolside's Laguna models. The always-current list with context sizes and benchmarks is on the OpenRouter page.
The openrouter/free router
Instead of a specific model you can request openrouter/free. OpenRouter then picks an available free model for each request. It is convenient for prototypes and fallbacks, but you do not control which model answers, so quality and output style can change between calls. For anything you test or ship, pin a concrete model ID.
Code
from openai import OpenAI
client = OpenAI(base_url="https://openrouter.ai/api/v1", api_key="sk-or-...")
reply = client.chat.completions.create(
model="nvidia/nemotron-3-super-120b-a12b:free",
messages=[{"role": "user", "content": "Write a haiku about rate limits."}],
)
print(reply.choices[0].message.content)How to get more out of the free tier
- Top up $10 once if you hit 50 requests a day — it multiplies the cap by 20.
- Handle 429 with backoff and a second model in the same family as a fallback.
- Add another provider. The same models are often free elsewhere: GLM and Kimi on NVIDIA Build, many routes on UnoRouter. Check the model card to see every free route.
- Batch small tasks into one request instead of many short ones.
For a broader overview of all 15 providers, read Free LLM APIs in 2026: the complete guide.