Official Free Tier AI API Routers

The category line is drawn at who operates the endpoint. Official free tier means the API belongs to the company that built (or directly sells) the models behind it: Google's AI Studio and Gemini API, Groq's inference endpoint, xAI's console, Cerebras's inference platform, Mistral's La Plateforme, the first-party endpoints of the model labs — and the big cloud platforms' free allowances. The distinction matters because the failure modes are different from relays and mirrors. An official tier is constrained by a published plan and a support process; a relay is constrained by whatever its operator decides on a Tuesday. When the official tier of a major lab tightens, it announces it on a pricing page you can re-read; when a relay tightens, you find out from your error log.

The 60 entries span the usual population: lab-direct APIs where the free tier is the on-ramp to a paid developer program, cloud-platform free allowances that ride on a broader cloud account, and a handful of first-party endpoints that are free as a permanent product decision rather than an acquisition funnel. All of them share one property this category exists to guarantee: the free access is documented by the operator, in writing, at a URL you can cite.

Three shapes cover almost everything in the slice. Trial balances give you a fixed spendable amount (a dollar value, or a token pool) that depletes and — for most labs — converts into a paid account if you keep going; Google's AI Studio pricing page and xAI's model pages are the reference examples, and the dataset normalizes the balance into the credit column. Permanent daily or monthly quotas give you less per period but never run out: Cloudflare's Workers AI free tier publishes per-model daily limits that "reset daily at 00:00 UTC", and Groq documents rate limits per model on its own console — these are the tiers that survive contact with a real project. And rate-capped free access, where the allowance is expressed as requests-per-minute or tokens-per-minute rather than dollars at all, which we normalize using the provider's published paid rates and flag when the assumption is ours.

The practical difference shows up in the credit column's volatility. Trial balances rank high (a $30 balance is $30) and then the entry goes quiet as balances expire. Quota tiers rank lower but their numbers hold steady across verification cycles — a stability signal that matters if you're choosing a long-running free home for a side project.

Free tiers systematically serve a narrower model list than the paid catalog. The frontier model is usually paywall-first; the free tier gets the previous generation, the small variants, or a subset of modalities. None of this is visible from the credit column, which is why the entry pages carry the tracked model list per provider — the honest question when comparing two official free tiers is not "who gives more dollars" but "does my workload's model exist on the free side, at a rate I can live with". The models directory (one page per tracked model, with per-provider pricing where known) is built for exactly that cross-check.

We track whether the tier exists, what its published limits are, when we last checked, and whether the endpoint answers our probes. We do not verify signup eligibility (student, region, one-account-per-person rules live in the provider's fine print and change), and we do not measure the free tier's rate limits under load — the published RPM/TPM numbers are the provider's, cited on the entry. Where a figure in the table is a normalization rather than a published dollar amount, it carries the flag. The provider's pricing page is the source of truth in every disagreement; the entry links it.

Run the same four questions against every shortlisted entry. One: quota shape — trial balance (spend it before it's gone), permanent quota (build on it), or rate cap (size your traffic to it); the credit note states which. Two: model fit — open the entry's model list and find the model your workload needs; a generous tier without that model is a different product. Three: the upgrade seam — official tiers are on-ramps by design, so check what happens when you exhaust the free side: does it flip to paid automatically (needs a card and a spend cap set before you start), error cleanly (safe to build against), or require a plan change (downtime you should schedule)? Four: the operator's history — the risk column and its rationale cover operator stability; among official tiers it mostly separates the majors with years of published terms from the young labs whose terms are still finding their shape.

The stacker's move, for completeness: official free tiers from different operators are independent allowances, so a client configured with two or three endpoints and a fallback chain gets a free capacity pool no single lab offers. The directory's compare tool and the per-entry OpenAI-compatible base URLs make that configuration a copy-paste job — and because each operator's revocation risk is independent, the pool is more stable than any member.