KIungefiltert

· 4 min read · Auf Deutsch lesen

Free Models on OpenRouter: Which Ones Work as Agents

Several free OpenRouter models handle tool calls well. With 50 requests a day, though, they're for experimenting, not for real work. And some hosts are allowed to train on your prompts.

AIsecurity

Translated from the German original with an AI language model, approved by me. Read the original

OpenRouter is a broker: one API gives you access to hundreds of language models from many providers. Some of them are free and carry the :free suffix. I wanted to know whether these free models work as agents. An agent is a model that doesn't just write text but calls tools on its own, such as reading a file, running a search or executing a script.

What I tested

On October 23, 2026, I checked whether the free models call tools correctly. This is known as tool calling: the model returns a structured call like "function X with parameter Y", and the agent software runs it.

Results from my test:

  • It worked with Nemotron 3 Ultra 550B (NVIDIA), Flint 1.0 (Recursive), Laguna S 2.1 (Poolside), Dots 3 (Dots Studio) and North Mini Code (Cohere).
  • Gemma 4 31B (Google) and Xing 4.0 only returned error 429.
  • Inkling (Thinking Machines) only ran for me inside agent tools.

These come from a single test run, not an independent benchmark. What I can back up: all of these models are listed as free, and according to their model pages and CostGoat's overview, they accept the tools parameter.

On error 429: according to the OpenRouter docs, it means "rate limit exceeded", so too many requests. I can't tell from outside whether the cause was an overloaded provider or a limit on my side.

On Inkling: the Inkling (free) model page doesn't say the free endpoint only works with agent tools. The only restriction it mentions is that the model isn't currently available through EU in-region routing. So the agent-only part is just what I saw, not something documented.

What has changed since the post

Not much. The CostGoat list was updated on October 29 and still shows all eight models as :free, each with tool support. It now lists 17 free models in total. That also shows how quickly the list changes. A model that's free today may be gone next week.

The limits

According to the official docs, confirmed by TrueFoundry, every :free model has these limits:

  • at most 20 requests per minute,
  • at most 50 requests per day if you've bought fewer than 10 credits in total,
  • 1,000 requests per day once you've bought at least 10 credits.

That's tight for agents. An agent often sends many requests per task, because every tool call is its own round trip, so 50 requests go fast. TrueFoundry also notes that failed requests can count against your daily allowance.

Privacy: the real price

"Free" doesn't mean you give nothing in return. According to TrueFoundry, 9 of the 19 free models they examined run on hosts whose terms allow training on prompts, including NVIDIA and Thinking Machines. Other hosts keep prompts for a while without training on them. Google AI Studio, for example, keeps them for 55 days. If OpenRouter can't determine a provider's policy, it marks the endpoint as both training and retaining, just to be safe.

That doesn't mean every prompt ends up in training data. But it's allowed. For an agent that reads files, sees code or parses logs, that's a real risk: whatever the agent sees goes to the provider.

What I recommend

  • Fine for experimenting: If you want to try out agent workflows or get to know a model, you get usable models at no cost.
  • No confidential data: No customer data, no credentials, no internal code. It's safest to assume this for every :free model.
  • Check your privacy settings: OpenRouter has privacy toggles and a data_collection: "deny" request option. That rules out providers that train on data, which often means losing free models too.
  • Don't depend on it: With 50 requests a day, 429 errors and a list that keeps changing, nothing in production should run on these endpoints.
  • Plan a fallback: If you use them anyway, set up a second model in case one disappears or gets throttled.

My verdict: good for playing around and learning, not for work you need to rely on.

Sources