OpenRouter Review: 458 AI Models, a Real Free Tier, No Subscription

▶ Listen to this article

This is an in-depth OpenRouter review born from real testing: one API key, 458 models, no subscription, and a free tier that actually works.

I have spent years juggling model APIs. Separate accounts at OpenAI, Anthropic, Google, then a half dozen smaller providers, each with its own dashboard, its own billing, its own rate limit table I was supposed to memorize. Every new model meant another signup. OpenRouter fixes exactly that, and it has a genuinely usable free tier that most people never find.

This is the review I wish I had before I hit the signup page. Real numbers from the live API, the free models that are actually worth your time, the rate limits that matter, and the honest cases where you should skip it.

What OpenRouter is

OpenRouter is a gateway. Not a model company. It sits in front of hundreds of open and proprietary models and serves them all through one OpenAI-compatible endpoint.

When you call a model, OpenRouter routes to the provider hosting it, and if that provider is down, the request falls over to the next one automatically. You get one key, one invoice, one analytics dashboard. The providers change underneath you. The interface does not.

The free tier is the real story

Here is what most review pages get wrong about OpenRouter. The free models are not demos. As of this month the API lists 19 models at $0 — 15 carrying a :free build, plus the free-model router, two Google Lyria audio previews, and inclusionAI’s free Ling 3.1 Flash — and the chat models among them are legitimate production models, not crippled trials.

The free lineup includes Google’s Gemma 4, NVIDIA’s Nemotron 3 and 3.5 series, Thinking Machines’ Inkling, and more. Four of them carry a full million-token context window — Nemotron 3 Ultra 550B, Nemotron 3.5 Lightning, Inkling and Inkling Small. A million tokens, free. One casualty of the last rotation: Z.AI’s GLM-5.2 lost its free build and is now paid only ($0.06 in / $7.00 out per million), so it is off this list. The roster keeps rotating: inclusionAI’s Ling 3.0 Flash also went paid, and its successor Ling 3.1 Flash is the free inclusionAI entry now. Qwen 3.8 27B lost its free build in this rotation too, and the stealth Space Bunny Alpha lane has left the catalog entirely. Two new free names replaced them: Apodex 1.1 Mini at 262K context and LiquidAI’s compact LFM2.5-2.6B at 65K.

Pick the right free model and you can run real workloads on it. For a prototype, a hobby project, or a low-traffic tool, the free tier alone is enough to launch.

The Free Models Are Not All Worth the Same

“Free on OpenRouter” describes the price, not the quality, and that list of free models hides a much wider spread than most people assume. The same independent measurement used across these provider reviews — the intelligence index and coding index from Artificial Analysis, plus the speed each model actually emits tokens at — puts the free lineup in order. Every model below costs $0 to call.

The spread is the story. The cheapest free model on this list and the best one differ by roughly six times on the same index, and both of them cost you nothing to call. That is why “just use a free model” is not a decision — the choice inside the free tier matters more than the choice between free and paid for most low-traffic work.

The best free model here is not the biggest name. InclusionAI Ling 3.1 Flash posts a 41.1 intelligence index with 215 tok/s, comfortably ahead of the entire Nemotron family and both Gemma sizes, in the same 262K context class. If you want one default for a free-tier prototype, that is the defensible pick. Second is Thinking Machines Inkling Small at 25.7 with a 52.9 coding score, and it carries the same full million-token context as its bigger sibling.

The largest models land mid-table. Nemotron 3 Ultra 550B is the biggest free option and scores 22.9; the 120B Super scores 12.8, below several models a fraction of its size. Parameter count does not buy position on this list — Gemma 4 26B also outranks its bigger 31B sibling again (16.7 against 14.7), the same result we found on the Google AI Studio free tier.

If you need speed rather than depth, Nemotron 3.5 Lightning is the fastest model in the table at 295 tok/s — more than six times the slowest here — but its 12.9 index is near the bottom. That is a fine trade for high-volume, simple chat and a bad one for anything requiring reasoning, which is the same split that shows up on every provider: speed and score rarely arrive together.

Three free lanes carry no independent row: Poolside Laguna XS 2.1, Laguna S 2.1 and Dots 3 Note Preview. Those are genuine coverage gaps rather than zeros, so nothing here should be read as a score for them.

Intelligence, coding and speed figures: Artificial Analysis, retrieved 10 October 2026. Context sizes are OpenRouter’s own listings, from the table above. The same index is used in our OpenCode Go review and Command Code pricing review, so scores are comparable across providers. Where Artificial Analysis publishes no measurement, this table shows a dash instead of a zero.

The rate limits you need to know

Free does not mean unlimited. Without credits, free models run at 20 requests a minute and 50 requests a day. That covers testing and building. It will not run a public chatbot.

Here is the trick that changes everything. Buy $10 of credits once, and the daily cap on free models jumps to 1,000 and stays there. The $10 is not consumed by the free models. It simply sits on your account for when you call a paid model. One purchase, and the free tier becomes genuinely usable for the week, not just the afternoon.

How the billing works

OpenRouter passes through whatever the model provider charges. There is no markup on the tokens. The company makes its money on the credit purchase itself, not on usage.

The cheap end of the catalog runs at fractions of a cent per thousand tokens:

A long running prototype can cost less than a coffee. When you only pay pennies for a thousand tokens, you stop thinking about usage and start thinking about what to build.

Bring your own key

If you already pay OpenAI, Anthropic, or Google directly, you do not need to buy OpenRouter credits at all. The BYOK program lets you plug your existing keys into OpenRouter and use it purely as a unified interface.

That setup gives you a million free routing requests a month. Past a million, OpenRouter charges a 5% routing fee on top of what the provider charges. For one interface over every key I own, with fallback and analytics included, I consider that fee a fair deal.

Privacy is handled, not promised

By default OpenRouter logs nothing of your prompts or completions. Not even when a request errors. Only billing metadata, timestamps, and token counts are kept. Your text stays yours.

There is an opt-in setting that trades that logging for a 1% discount on usage costs. It is off by default, and you have to switch it on deliberately. For a company whose whole business is passing data between you and model providers, that default deserves credit.

Those model suffixes explained

The model IDs end in a suffix that tells OpenRouter how to route the call. It looks like noise until you know what they mean.

• :free uses free tier hosting where available.
• :nitro routes to the fastest provider.
• :floor routes to the cheapest provider.
• :thinking enables reasoning mode on models that support it.
• :extended routes to a provider offering a longer context window.

So deepseek-r1:free gets you a reasoning model for nothing, and llama-3.3-70b:floor gets you a paid model at its cheapest source. The suffixes stack on top of any model you already know.

Where it falls short

I tested the free tier properly, and I want to be honest about the limits.

Do not run production traffic on free models. No SLA, hosting can drop out, and 20 requests a minute dies the moment a second user opens your app. The free tier is for building and proving an idea, not for scale.

The free roster also rotates. A model that is free today can lose its free hosting next month. The :free router picks whatever is available, but you lose the choice of a specific model.

And the credit purchase fee is the catch. If you spend heavily on a single provider, the direct route is sometimes cheaper. OpenRouter wins on convenience and breadth. It does not always win on pure price.

Who should use it

Get it if you compare models, if you build something that might switch providers, or if you want to prototype on free models before paying. Get it if you already hold several provider keys and want one dashboard.

Skip it if you use one model from one provider and nothing else. The credit fee is pure overhead there, and you are better off going direct.

FAQ

Do I need a credit card to start?

No. The 19 free models work the moment you sign up. Credits are only needed to lift the daily cap or call paid models.

Is it a drop in replacement for OpenAI?

Mostly. Point your client at the base URL, drop in your OpenRouter key, and the request format matches. Code that targets OpenAI generally works unchanged.

What does OpenRouter charge?

A fee on credit purchases, and a 5% routing fee on usage above a million BYOK requests a month. Model pricing passes through with no markup.

Which free model is best right now?

By the measured scores in the table above, inclusionAI’s Ling 3.1 Flash leads the free group at a 41.1 intelligence index and 215 tok/s. Thinking Machines’ Inkling Small is the pick when you want the million-token context, and Cohere North Mini Code or the Poolside Laguna pair are the code-focused options.

Can this replace my provider?

If you use several, yes, it replaces most of the friction. If you use one, it is overhead. It is a gateway, not a model.

The bottom line

The thing I keep coming back to is the friction it removes. No new account at every lab. No memorizing five rate limit tables. No logging into a different billing page to understand a single invoice.

One key, and behind it, everything. Start on the free tier, put $10 down if the daily cap bites, and decide from there. If it is not for you, testing it costs nothing at all.

Sign up at openrouter.ai. No credit card needed to start.