free_premium_ai: a curated list of free AI access platforms

A directory of free quota, not a free AI service

free_premium_ai is best understood as a maintained list. The repository description calls it a collection of platforms where a fairly capable AI can be used for free, and it says outright that the author tries to pick stable platforms while warning about the impossible triangle: free, stable and high performance cannot all be had at the same time. The same passage notes that free credit usually exists to support developer testing and to attract paying customers, and that free access costs more of your time.

So nothing here is a model, an endpoint or a proxy. It is a set of curated entries, each naming a platform, the advanced models that platform exposes, and the terms under which the free quota is granted. The audience is developers looking for somewhere to run a capable model without opening a paid account, and the author explicitly says the project does not recommend any platform and that a platform which dies should be reported through an issue or a pull request.

The repository is a JavaScript project with a small structure: index.html, build.js, a data directory, a scripts directory, a tests directory and an assets folder. That layout says the README and the website are generated from one dataset, so the two never drift apart.

How a model qualifies for the list

The inclusion criteria are specific enough to be checked, which is unusual for a list like this and is the main reason to trust the entries. The first requirement is model capability: a model must score above 70 on the Global Average of LiveBench, or in the author's words it must be at least as capable as GLM5.2. The README notes the leaderboard updates slowly, so while looking for a substitute the bar is temporarily set at that level.

Where a model has no API, the check runs through a prompt based detection method, and where an API exists the author points at a detector service for it. The README also says that capability reduced by system prompts or by restricted features, for example tool calling, does not disqualify a model, but the defect is noted in the entry.

Models pretending to be something else are excluded outright, and the example given is a GLM5 presented as a Claude model. For open source models the README says there is no effective test method yet and invites submissions of one. That is an honest statement of a limit, and it means an open source entry in the list rests on the submitter's word.

The two verification services the criteria lean on

Two external services carry much of the verification weight, and the README links both, so a reader can repeat the check rather than take the verdict on trust.

https://livebench.ai/
https://hvoy.ai/

LiveBench is the benchmark the capability threshold is measured against, and it is also a usable alternative to the list for anyone whose real question is which model is strongest rather than where it is free. The difference in approach is worth spelling out: LiveBench ranks capability across models and says nothing about quota or registration friction, whereas the list inverts that and treats capability as a filter to be passed before the pricing and access details matter.

The detector service is the second tool. It exists because a provider naming a model in its own documentation is not evidence, and the README is blunt about the failure it catches. For a reader deciding whether to spend an afternoon wiring up a new platform, running the check is a cheaper test than the trial the entry describes.

What the free quota and registration rules have to look like

The second inclusion criterion covers the free allowance itself, and it is split into three acceptable shapes. A platform may give a large balance at registration, in which case strict registration is allowed, for example real name verification or a phone number binding, and the entry then says so under the quota and registration note. A platform may grant a daily or weekly check-in allowance, or hold free access to advanced models over time, again with strict registration allowed. A third shape is a small registration balance with loose registration, which the README describes in deliberately indirect terms.

The third criterion restricts what kind of platform is listed. Only paid platforms are collected proactively. Public benefit sites are not collected proactively, though a site owner may submit one directly with a statement of the rules, and submissions from anyone other than the owner are not accepted. That is a deliberate editorial line, and it keeps the list pointed at services with an incentive to keep running.

The fourth criterion covers how the platform may be called. No protocol convention is required and private protocols are allowed, with the author writing a converter from a private protocol to the OpenAI format before listing it. Platforms that only work inside a programming tool are acceptable too. The practical effect is that some entries will not work with a plain HTTP client until a converter exists.

Sorting by expiry and the affiliate policy

The sort order is the most useful idea in the project and it is explained in one line: channels that expire easily go first, and everything else follows in the order it was added. A reader therefore sees the fragile entries where they are most likely to be wrong, which is the opposite of what a list sorted by popularity would do.

Submissions run through pull requests or issues, and the project publishes a fixed template. Each entry states the platform name, the registration link, the advanced models that meet the criteria, and a note on the free quota and the registration restrictions, with a screenshot of test results attached. The website offers a submission form that opens a prefilled GitHub issue, which removes the need to know the Markdown layout.

Affiliate links are allowed and disclosed. The stated rule is that if a reader is the first to find a platform and submit it, the affiliate link stays with that reader permanently. The README also says the project itself does not recommend any platform, and asks that an operator who wants their service removed open an issue with proof of domain ownership.

That combination, transparent commissions plus a removal right, is the part of the design that makes the list tolerable rather than merely promotional.

The website is the more usable copy of the same data

The README points readers to a hosted site and recommends it, describing a card based layout, full text search and light and dark modes, with content kept in sync with the document. The project name in the repository, free_premium_ai, matches the site address.

https://free-premium-ai.pages.dev

For the actual entries, the site is the better surface. A markdown list is awkward to scan when you are looking for one provider among thirty, and the entries differ along the axes a reader cares about: the free balance, whether a phone number or real name is needed, whether an API is offered, and which models are named. The card layout and the search field address exactly that, and the theme toggle matters for a page that gets opened repeatedly.

The repository structure backs this up. build.js and the data directory suggest the site is generated from a dataset rather than hand maintained, with scripts and tests alongside. That is the right way to run a list this long, since the same source can feed both the README and the site without one drifting behind the other.

One caveat visible from the entries themselves: several describe services that are only reachable through an official channel with no API at all, so a reader looking for a programmatic endpoint has to read the note rather than the model name.

Where this list will mislead you

The first limitation is the one the project states for itself. Free quota exists to attract paying customers, so the amounts and the daily allowances get cut. One entry in the list says its allowance was cut substantially compared with earlier, and another notes that a daily token figure is now the main limit. Nothing in the repository can prevent a provider from reducing a grant between one read of the list and the next, and the sort order only tells you where fragility is likely, not whether a specific entry still works.

The second is verification depth. The capability bar is a LiveBench threshold, which measures general ability on the leaderboard, not the behaviour of one provider's deployment of that model. Entries can also be degraded by a provider's system prompt or by restricted features, which the criteria accept and merely note.

The third is model identity for open source entries, where the README admits there is no effective test method and relies on submitters. That is where a plausible looking but wrong entry is most likely to sit, and it is also the case a careful reader cannot check with the tools the project itself recommends.

The fourth is the affiliate economics. Commissions are disclosed and the submission policy is fair, but a link that pays the submitter still shapes which platforms a reader reaches first, particularly in a list ordered by how easily a channel dies rather than by quality.

The fifth is coverage. Only paid platforms are collected, and only when someone bothers to submit one, so the absence of a platform from the list means nothing.

Licence, releases and who should read it

The licence is MIT, and the repository ships a LICENSE file, so the list data and the site generator can be reused and modified under permissive terms as long as the notice travels with them.

The project is not archived and the last push was on 2026-10-16, so the list is being worked on. The recent releases are tagged by date rather than by version number, with tags such as 2026-8-28, 2026-05-16 and 2026-05-15, which reads like a publishing cadence rather than a software release history. The README also records that the account was wrongly suspended by GitHub at one point and later restored, with updates resumed and the site recommended for bookmarking.

None of that makes the list a dependency. It is a reading resource, refreshed on a schedule, and the right way to use it is to treat each entry as a lead rather than a promise: open the platform's own page, check the quota and the registration requirement listed in the entry, and decide whether the free tier is worth the phone number binding it asks for. The value of the repository is that it tells you which of those leads are worth the walk, and it says plainly that the free, stable and fast combination is not on offer.

Editorial conclusion

free_premium_ai is worth reading if you want to find a paid provider giving away usable model quota, and you can accept that the entries expire, the quotas shrink and several links carry affiliate commissions. It is not a way to obtain models that were never meant to be free, and it makes no claim that any listed platform will still work. Before relying on a specific entry, open that platform's own page and check the quota and registration rules listed there, and read the inclusion criteria in the repository rather than assuming a listed model is genuinely what its name says.

Frequently asked questions

Official sources

1. License: MIT
2. Project website
3. README
4. Releases
5. swufe-xiongmin/free_premium_ai on GitHub