OpenRouter Free Models: What's Real, What Disappears

OpenRouter lists dozens of models at $0. Some are permanent, subsidised by the labs that make them. Others are previews that vanish with no notice.

Both look identical in the model list. Here is how to tell them apart — and why it matters if you are building anything on top.

The three kinds of free

1. Genuinely free tiers of real models. A lab publishes a :free endpoint alongside its paid one — usually the same model, sometimes with tighter rate limits. NVIDIA's Nemotron family does this. These are stable: the paid version anchors the free one, and the lab has a reason to keep both running.

2. Free-while-in-preview. A new model launches at $0 to gather usage and feedback, then gets a price. The model itself survives; the free access does not.

3. Anonymous stealth models. No named developer, no published pricing beyond $0, no commitment. These are the ones that disappear.

The third category is where people get burned, because it is also where the most exciting models often show up.

A worked example: Ox Alpha

On 20 August 2026, a model called stealth/ox-alpha appeared on OpenRouter. Free. A 1M-token context window. Strong at coding and long agentic tasks. OpenRouter stated plainly that it was not the developer and could not say who was.

Developers piled in. It was, by most accounts, genuinely good.

We were asked whether to add it to BrahmAI's free tier. We said no, and wrote down four reasons:

  1. Anonymous provider — no accountability if it broke or vanished
  2. $0 is a preview promo, not a rate — it would reprice or disappear without notice
  3. Unconfirmed identity — community fingerprinting suggested a disguised Z.ai model
  4. Free tier is the worst place for something unstable — highest volume, most cost-sensitive users

Within a week it was gone from the catalogue.

Z.ai later confirmed what the fingerprinting had suggested: Ox Alpha was an anonymous preview of GLM-5.3-Flash. It did not die — it graduated, and relaunched under its real name with a real price.

Anyone who had wired it in as a default was left with a model id that no longer resolved.

How to tell the difference before you commit

Ask four questions of any free model:

Who made it? If the answer is "nobody will say", treat the listing as temporary. A named lab has a reputation to protect and usually a migration path.

Is there a paid version of the same model? A :free endpoint sitting next to a priced one is a much safer bet than a model that exists only at $0. The paid tier is what keeps the lights on.

Is $0 described as a price, or as a promotion? Preview pricing is usually flagged as such somewhere. If it is not flagged at all and the provider is anonymous, assume promotional.

What breaks if it disappears tomorrow? This is the real test. A free model you experiment with costs nothing when it goes. A free model wired in as a default breaks every user who depends on it.

What we actually ship

We evaluate models across the industry for BrahmAI, so this is not a hypothetical for us.

Our rule is simple: a model needs a real name, a real price, and a real owner before it goes in front of users. Ox Alpha failed all three. GLM-5.3 — the same family, once it launched properly — passed all three, and shipped.

Our free tier runs on models that pass that test — current releases from named companies like NVIDIA, Google, DeepSeek and Alibaba, each of which stands behind the model and keeps it running. If the terms on any of them ever change, the model still exists and we handle it. Nothing breaks for the person using it.

That is a boring answer. It is also why the free tier still works the same way it did last month.

The rate limit question

The other thing worth knowing about free models on OpenRouter: free endpoints are rate limited more aggressively than paid ones, and those limits are shared across everyone using them.

In practice this means a free model can return a 429 error at busy times through no fault of your own. If you are building something that needs to work reliably, either handle that gracefully or use a paid endpoint.

However you build, the important thing is that a busy provider degrades visibly rather than hanging. A feature that says "temporarily unavailable" is far better than one that spins forever while the user waits.

The short version

Free models on OpenRouter are worth using. Just know which kind you have.

A :free endpoint from a named lab, sitting beside a paid version, is something you can build on. An anonymous model at $0 with no published owner is something to experiment with and nothing more.

The exciting one is usually the second kind. That is exactly why it is worth checking before you depend on it.