Free AI Router

A free ai router is an AI API router with no paid tier at all: the only way to use it is the free access it offers, and there is nothing to buy once that access runs out. In the UPROUTER dataset this is the routerType "free-only" — 39 published entries as of the last check, out of 388 tracked routers. The other three router types exist for contrast: freemium (38 free tiers plus paid options across the directory), paid-only, and enterprise-first offerings.

A router is a service that exposes AI model APIs — usually an OpenAI-compatible /chat/completions endpoint — so one key reaches models that otherwise sit behind several vendor consoles. Free-only routers are interesting precisely because they remove the billing surface entirely: no card on file, no spend limit to configure, no invoice. What you give up is the exit ramp. A freemium router lets you start free and pay as you grow; a free ai router caps out where its free access ends.

The 39 entries fall into recognizable shapes. Some are platforms that are free as a business model (a shared endpoint anyone can call, or a free credit grant for trying the product). Others are community projects that relay upstream inference. The access column in the sample below tells you which: "API key" means you register and get a key that encodes your free allowance; "No auth" means the endpoint is callable without any credential at all — convenient, and also why operators can revoke it unilaterally at any moment.

The absence of a paid tier is not just a price point; it changes the risk profile. With a freemium provider you are a customer: terms of service apply to a defined account, support channels exist, and the provider has a financial reason to keep the free tier roughly as documented (it is their acquisition funnel). With a free-only router you are a guest. The access can be rationed, tightened or switched off by an operator decision, and your code will only find out by failing.

That makes free-only routers a genuinely good fit for specific jobs and a bad one for others. Good fit: prototypes where you are learning the model's behavior, local development, low-volume automation (a nightly job, a small personal tool), and education, where the point is wiring up the call. Bad fit: anything user-facing or time-bound, anything where a silent rate-limit change at 9am costs you money, and anything that needs an SLA — no free-only service in this slice offers one, because there is no paid side to attach one to.

There is also a data-handling angle: a no-auth relay cannot hold you to a data policy, and your prompts may sit in shared infrastructure you cannot audit. For sensitive inputs, prefer "API key" entries with a named operator, and read the risk rationale first.

Three mechanisms cover nearly everything in this slice, and they fail differently, so it is worth knowing which one a given entry uses — the entry page records it in the free-credit note with the source URL.

One-time grants. The operator credits your account with a fixed allowance at signup, and the balance depletes. DeepSeek is the cleanest example: its pricing page states that fees are deducted from a topped-up balance "with a preference for using the granted balance first when both balances are available" — the granted balance is the free part, and once it is gone, the account is a paying account or nothing. On a free-only router the second option is the only one, so the grant is the whole relationship.

Recurring quotas. Some platforms reset free usage on a schedule rather than a balance. Cloudflare's Workers AI free tier, for example, publishes explicit daily limits that "reset daily at 00:00 UTC"; the dataset's free-credit value for such entries is a normalization of the quota at the provider's own pay-as-you-go rate, which is why the figure is an estimate and flagged where the assumption was ours.

Shared endpoints. The router serves inference from pooled capacity and meters nobody individually. Hugging Face's Inference Providers route a single token to eighteen partner providers and state "no extra markup on provider rates" — that is the aggregator model applied to free access; Pollinations runs the same shape (one OpenAI-compatible API for text, image, audio, video) with "free Pollen for prototypes & testing". Shared capacity is the most convenient and the most revocable: no contract protects you when the operator changes their mind.

Every entry in this directory carries a community-evidence risk rating (low / medium / high / unrated) with the reasoning on the entry page, but the rating matters most on this slice. The reason: the free-only business model attracts relays that pool API keys obtained through other people's accounts or through upstream free tiers, and that pattern sits in direct tension with upstream providers' terms. OpenAI's usage policies prohibit sharing accounts or reselling access to the API; a relay that fronts shared keys is structurally exposed to being cut off, and your application inherits that exposure.

The dataset does not try to hide this distinction. Entries whose operator identity, terms or funding are unclear tend to carry "medium" or "high" with a rationale naming the specific concern (young service, vague terms, unclear data handling, ToS friction — the rubric is on each entry and summarized on /about). "Unrated" means the evidence is not there yet, not that the entry is clean. Treat a high-risk free relay as a convenience endpoint you can lose at any time, and design around that rather than discovering it.

To keep this page honest about its own limits: UPROUTER runs best-effort status probes against each endpoint (the "status" column and the probe history on each entry), normalizes published free-credit figures into comparable USD estimates and flags every estimate , and records the date each figure was last checked. We do not benchmark latency or throughput, do not guarantee uptime, and do not re-verify each provider's terms — the ToS claims link to the providers' own policy pages, which change. Free tiers are the most volatile part of any AI pricing table; when in doubt, the provider's pricing page is the source of truth, and the entry page links it.