How to choose a safer third-party AI aggregator or mirror: a checklist

“A site where you can use several models” is not automatically safe. Whether it is safe depends on whether it wants your official passwords, who receives the money, whether the terms can be found in public, and whether the model it claims to offer actually matches the name. Below is a checklist you can tick through, item by item, to decide whether a third-party aggregator, mirror, or relay is worth handing conversations and payments to.

This checklist does not target any specific gray-market site, and it does not teach anyone how to use a mirror of unknown origin. The only goal is that you can finish the checks yourself before you pay or upload a file.

Why “it works” is not enough

A third-party platform usually works like this: you sign in and pay on its website, send a question to its servers, and it forwards the request to one or more model providers, then shows you the answer. That means:

• Your account and money go first to this operator, not to OpenAI or Anthropic.
• Your prompts, attachments, and generated results at least pass through the operator, and may then reach the model provider.
• Memory, after-sales support, and data-control options from official products do not appear on this page by default.

So the standard should not only be “can it open ChatGPT / Claude right now,” but: if something goes wrong, can you find a party that is accountable, will credentials be taken, and will content be used in a way you cannot accept.

For background on payments and accounts, see How to use Claude, ChatGPT, and Grok from China. For details on uploading files, see Is it safe to upload files or PDFs to an AI?.

Checklist (go through these before you pay)

Check in this order. If any of the first three items is clearly a fail, you do not need to try further.

1. Does it want your official password or login credentials?

This is the hardest line.

• The safer practice: register only on that platform’s own site, and sign in with that platform’s own login. It should not ask for your password on the official ChatGPT, Claude, Gemini, or Grok websites, for SMS or email codes, or for a complete login cookie / session token.
• A practice to treat as a warning: it asks you to “sign in officially first, then send over a piece of session information,” “send the verification code to support,” or “hand the account and password to someone else to open the plan.” Once that flow is complete, the other party may be able to act as you on the official account.
• What not to do: turn off two-factor authentication in order to start a membership, or forward a verification code to a stranger in a chat app.

An independent product and “operating an official account on your behalf” are completely different paths. If the former gets worse, the loss usually stops at the money and data on that platform. The latter can mean losing the official account itself.

2. Is the domain or the page impersonating an official product?

When you open a site, look at the address bar first, not only the logo and colors.

• Check whether the domain matches the platform you intended to find, whether a letter has been added, whether a similar suffix has been swapped in, or whether an official brand name has been assembled with words such as “chat,” “login,” or “plus.”
• Official product domains are public, for example chatgpt.com, claude.ai, gemini.google.com, and grok.com. A third-party platform should use its own brand domain, and should state in its terms that it is an independent service.
• Browser certificate issues, an unusual number of redirects, or a full-screen popup that urges you to enter an official account as soon as you arrive are all reasons to stop.
• If a page copies an official interface extensively without saying “this is not an official site,” treat that as a warning too.

The point of a look-alike domain is often to get you to hand over official credentials in a login box you already recognize. Even if you can chat a little afterward, that starting point is already unsafe.

3. Are the terms, privacy policy, and refund rules publicly visible?

Before you pay, you should at least be able to open these pages without signing in, and see:

• The operating entity. Which company or organization provides the service, rather than only a handle or an instant-messaging account.
• A privacy policy. What information is collected, whether it is shared with providers, and whether you can access or delete your data. If conversations and uploaded files are not mentioned at all, at least treat that gap as a risk, rather than assuming “it must not be saved.”
• Terms of use. What the service is and is not, whether it has an affiliation or authorization relationship with the model vendor, and which uses are prohibited.
• Refund and cancellation rules. How to cancel, whether renewal is automatic, how long you have to request a refund, and how long a review takes.

If these pages do not exist, are too thin to act on, or can only be “sent privately by support after you pay,” the rights and duties are unclear. When something goes wrong, it is hard to prove what you bought.

Take Trymodel as an example (checked on 11 November 2026, for comparison, not as a claim that “only this one passes”):

• The operating entity is stated in the terms and conditions: operated by TUCANTELA SL, and it states that it is not affiliated with OpenAI, Anthropic, Google, xAI, or similar vendors, and does not claim endorsement or approval.
• The privacy policy is publicly accessible. It describes the handling of account information, usage data, device information, cookies, and payment information, and it states that personal information is not sold and that full card numbers are not stored on the servers; it does not clearly state the retention period for conversation content and uploaded files, or whether they are used for training. That gap should be read as it is; see the article on upload safety.
• The refund and cancellation policy states how to cancel, that you should cancel 72 hours in advance to avoid the next cycle’s charge, that you can request a refund review within 30 days, and the limited review condition when there are fewer than 5 messages in the current cycle.

Clear terms do not mean zero risk; but if the terms cannot be found, the risk can barely be assessed at all.

4. Can the payment method be checked, and will you be charged without knowing?

Inspect the payment page, not only the marketing copy:

• Can you see the amount due this time, the currency, and whether it is one-time or auto-renewing.
• Is the receiving method a channel the site has published, rather than a transfer to a personal account.
• If there is auto-renewal, are the cancellation path and the cutoff time stated.

The Trymodel pricing page (plan prices in this English version follow the list checked on 14 November 2026) states that the weekly, monthly, and yearly passes and credit packs then on the page can all be purchased as a one-time payment, paid with cards or Alipay, and do not renew automatically when they end. WeChat Pay is available when you check out in Chinese yuan on the Simplified Chinese site. If you want auto-renewal, monthly and yearly passes can choose “auto-renewal (cards only).” A quarterly pass was not shown that day. The refund policy requires that, to avoid the next cycle’s charge, you submit a cancellation at least 72 hours before the next billing date. The full figures and rules are in Trymodel pricing and billing.

If a site only says “we will not charge extra,” but you cannot find a cancellation path or an order record, do not pay on a verbal promise alone.

5. Use the same question to check whether the model matches the name

A flagship model name on the label does not guarantee that the backend is actually calling it. Before you invest more time and files, run one comparison with the same question:

1. Prepare two or three questions you know well and can judge easily. For example: a clear bug in code you wrote, a question that needs a recent public fact, or a summary of a short text you have read. Avoid uploading real business files at the start.
2. On that platform, choose the model it labels, and save the full answer.
3. If you also have an official entry point, ask the same question, with the same source text, in the official interface, and compare whether the reasoning steps, citations, and obvious errors are at a similar level. If you have no official entry point, you can also ask the same question of another model on that platform, and see whether the differences are reasonable: if several “completely different” models produce almost word-for-word the same output, or if a top-tier model and a free small model show no distinguishable difference, you should trust the label less.
4. See whether the interface shows the model’s full name, whether you can switch mid-thread, and whether the answer inflates the model name and its capabilities clearly beyond public information.
5. Treat this comparison only as a check on “whether the label is roughly credible.” Do not write it up as a formal evaluation, and do not invent scores.

A mismatch is not always malicious. The model may have been replaced after it went offline, the name may lag, or routing may be wrong. For the user, the result is the same: you paid for a flagship model and do not know what you are actually using. If it does not match, stop uploading more material, and keep the conversation record so you can ask support.

6. Decide first what you will not upload

However orderly a platform looks, the following should not be casually handed to a third party, and should not be pasted into an official free chat box either:

• ID numbers, bank card numbers, account passwords, API keys, internal network addresses, and unpublished access credentials.
• Unauthorized customer lists, employee compensation, medical records, information about minors, or other regulated data.
• Contracts that are not yet public, secrets in source code, or unpublished financial data.
• Files that someone else allowed only you to read, and did not authorize for further distribution.

Trymodel’s terms also state: do not submit confidential, regulated, or highly sensitive information without authorization. More specific redaction methods are in Is it safe to upload files or PDFs to an AI?.

A practical rule: if this page appearing on a public website would make you want to take it down immediately, do not upload it “just to try.”

7. Start with a small purchase, then decide whether to spend more

Even if the earlier items all pass, it is still better to:

• Buy the shortest or lower-priced plan first, and use it for a few days on real but publicly discussable tasks, rather than paying for a year up front.
• Immediately after payment, check the amount due, how it was received, and whether auto-renewal was turned on.
• See whether usage is in the range you expected; many platforms (including Trymodel’s terms) state that reasonable-use or operational limits exist, and do not guarantee a fixed number of messages.
• During a trial, do not upload sensitive files, and do not make that platform your only backup.

On the English pricing page checked on 2026-11-14, the weekly pass is $3.99 / 7 days, purchased as a one-time payment that does not renew automatically when it ends, and can be paid with a card or Alipay (WeChat Pay when you check out in yuan on the Simplified Chinese site). That is suitable for completing the comparison above, not for proving that “you can already entrust all of your work to it long term.” If you renew after a trial, see Trymodel pricing and billing.

Apply the checklist to the site you are looking at

You can score yourself with a short table. If any item is “no” or “cannot be seen,” do not add a payment or an upload yet:

Trymodel’s chat page, pricing page, privacy policy, and refund policy can serve as samples of “whether the pages are public.” When you compare, open the original pages; do not rely only on excerpts in this article. The policies took effect on 7 May 2026 (privacy policy) and 17 October 2026 (terms and refund policy). If they are updated later, use the new pages.

If you also want to compare the shape of different aggregator tools (in-thread switching, a points marketplace, side-by-side windows, token routing), see How Trymodel differs from Poe, Monica, ChatHub, and OpenRouter. To check which models are on the site that day, see New AI models: which to use.

FAQ

What is the difference between a third-party aggregator and a “mirror site”?

There is no single industry definition. In practice, “aggregator” often means one interface connected to models from several vendors; “mirror” is often used for a site that imitates an official interface, or even makes it look as if you are using the official site. A more useful distinction for users is: is it an independent product, does it want official passwords, and are the terms public. A name that sounds legitimate cannot replace these checks.

If it does not ask for official passwords, is it necessarily safe?

No. Not asking for official passwords only means it is less likely to take over your official account directly, but conversations may still be stored, forwarded, or used in ways you do not understand, and payments may not match the marketing. Passwords are only the first item on the list, not the only one.

How can I tell whether it is actually using the model I selected?

The most realistic method is: use questions you know well, run the same question on that platform and on an official entry point (if you have one), and watch whether different models show reasonable differences. This is still a judgment from experience, not laboratory-grade proof. If it does not match, reduce how deeply you use it, and do not keep uploading sensitive material.

Why start with a small trial instead of buying a yearly pass at once?

Because features and billing can only be checked after you have paid once and actually used the product. A small trial keeps the loss in a controllable range, and also lets you confirm whether auto-renewal was turned on and whether usage matches what you expected. A longer plan should come after a trial.

If I have found the terms, does that mean I can upload internal files?

No. Public terms only mean the rights and duties are relatively clear. Once a file leaves your device, it enters the operator’s systems and possibly the model provider’s. Even if the platform allows uploads, content that should not be shared still should not be shared.

If something goes wrong, what should I keep?

Keep the order number, payment receipt, the plan description at the time of purchase (or a screenshot of the page), correspondence, and conversation excerpts that do not contain passwords or ID numbers. Communicate first through the contact method the site has published. If a bank or payment dispute is involved, follow the payment institution’s rules, and avoid sending more sensitive information through the same channel.

FAQPage structured data

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "When choosing a third-party AI aggregator or mirror, what is the most important safety check?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "First see whether it wants your official ChatGPT, Claude, or Grok password, verification code, or login credentials; then check whether the domain impersonates an official site, whether the terms and refund policy can be read in public before you pay, and whether payment will renew automatically. Do not sign in on a site that asks for official credentials."
      }
    },
    {
      "@type": "Question",
      "name": "How can I tell whether the model a third-party platform claims to offer actually matches the name?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Use two or three questions you know well, choose the model the platform labels, and, if you have an official entry point, run the same question there for comparison. You can also compare whether differences among models on that platform look reasonable. This is only a judgment from experience, not a substitute for a formal evaluation. If there is an obvious mismatch, stop uploading more material."
      }
    },
    {
      "@type": "Question",
      "name": "Why should I not upload internal files to a third-party AI platform at the outset?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Prompts and attachments usually pass through the platform operator’s servers, and may then be sent to the model provider. Once ID documents, secrets, unauthorized customer data, or confidential contracts are uploaded, they are outside the range you can control on your own. Start with a small order and with tasks you can discuss in public."
      }
    }
  ]
}

Facts were checked on 11 November 2026 (Asia/Shanghai). Trymodel plan prices in this English version follow the list checked on 14 November 2026. Excerpts of terms and pricing in the checklist should be read against the original pages that day.