Frequently asked questions

No. Type your question into the input field and the chat starts right away. A guest record is created by itself on your first request, with no form at all: no email, no name, no SMS confirmation. A guest and a registered user get the same allowance, so signing up does not buy you messages.

Nothing. There are no paid plans at the moment: payments are not connected, and Premium cannot be switched on even by pressing the button — the service answers "Payments are not connected yet". No financial obligations arise between you and the service.

Access is limited by quotas rather than by money: a monthly allowance of requests to the models, the size of your storage, the number of files and the number of images per day. The "Premium" item in the menu shows what the subscription will include once it exists. We are not naming a date.

One thing, and it matters: access to your chats from somewhere other than this browser. A guest is recognised by a random token stored in the browser and by nothing else. An email and a password replace that token — you can sign in from your phone or another computer, and come back after clearing the browser. The "Save history" button on the Account page leads exactly there.

Nothing is lost in the process: the guest record becomes your account, and the chats and files stay where they were.

No. There is no app; the service opens in a browser, on a computer and on a phone alike.

No. Your browser only talks to easygpt.app; it is the service's server, not you, that reaches OpenAI, Anthropic, Google, xAI and the rest. So no VPN is needed, and neither is a foreign card or a foreign phone number — nor, for that matter, an account.

An internal unit of account. An exchange with a model costs more credits the more that model costs at the provider, which is why the allowance is counted in credits and not in messages: the same number of messages costs differently on a fast model and on a strong one. Credits cannot be bought, exchanged or passed on — they are a spending counter, not money.

The allowance is granted on your first request to a model and renews every thirty days; the date of the next renewal is shown on the Account page. Whatever you did not spend does not carry over into the new period.

The service deliberately does not show that counter: a number over your head hurries you and distracts from the task. You find out you have run out at the moment you hit the limit — a message arrives with the figures, how much is needed and how much is available. The rest is on the Account page: storage used and free, and the date the allowance renews.

The service does not lock up. First it quietly moves the request to the fastest model, and most of the time you will not even notice. If even that is out of reach, you get a refusal: "Not enough credits for this request". Then the options are to wait for the new period — or to write shorter: the price of an exchange includes both the length of the request and the length of the answer.

Everyday questions are handled by the fast ChatGPT-5.6 Luna: it is the default in the list and it answers everyone, guests included. The others are visible in the list too — DeepSeek V4 Pro (works through a task step by step: sums, logic, code), Grok 4.5 (the latest news and social media), Claude Sonnet 5 (writing and editing), Gemini 3.1 Pro (going through documents and images), Sonar Pro Search (searches the web and gives links to its sources), Claude Opus 5 (the strongest).

All of them except the first are marked "Premium", and Premium cannot be switched on right now. So the fast model is the one answering you, and the other rows show what the service has — not what you can use today.

There are two reasons, and they look the same in the list. The first: the model costs money at the provider, and the free path only allows the ones that cost the service nothing. The second: it is a reasoning model — its thinking is billed like an ordinary answer and noticeably delays the first word, so those are closed off on the free path separately from price.

Because it produces plausible text rather than checked facts, and a confident tone in an answer means nothing at all. Invented dates, quotations, statute numbers and links are ordinary behaviour for any language model, not a fault in the service. Nobody checks or corrects the answers: it says so plainly under the input field. Verify anything that matters against the original source.

That is the model developer's decision, not the service's: each of them has its own rules, and they fire before the answer ever reaches us. Saying the same thing in different words and removing whatever the rule catches on usually helps. The service cannot appeal such a refusal or work around it.

What goes to the model is not the whole conversation but its last messages: a request has a budget, and everything that did not fit is cut off. That is why the beginning is forgotten in a long chat — it is the edge of the budget, not a memory failure. If the conversation has drifted far, starting a new chat and repeating what matters in it is both cheaper and more accurate.

Yes, with the paperclip in the input row, but the file has to be text: .txt, .md, .csv, .json, .log and source code (.py, .js, .ts, .html, .css, .xml, .yml) all work. The file is read in the browser itself: it never reaches the server, its text is dropped into the input field, and you see exactly what will go to the model. The ceiling is 300 KB; anything larger has to be pasted in as the part you need.

PDFs, .docx and spreadsheets are not read by the paperclip: there is no plain text inside them, and the request would fill up with control characters. Save such a document as text, or copy the part you need by hand.

In the Files section — the same place generated images land in. The storage is private, you can make any folders you like, and the file type is not restricted. The limit on the free path is 10 MB per file and 50 MB for the whole storage; the space used and left is shown both in Files and on the Account page. Files live separately from the chats and do not disappear along with a deleted conversation.

On the Images page: the "Create" tab draws from a description, the "Change" tab edits a picture of yours ("make the background white", "remove the people at the back"). You can pick a style and a frame, and the result lands in Files.

Today, though, drawing will not work: image generation is part of Premium, and Premium cannot be switched on. This is not greed — every image costs the service real money on every request, so this one thing cannot be free. Uploading your own images, keeping them and downloading them works without Premium.

There is: the microphone in the input row. Speech is recognised by the browser itself, not by the service — no audio is sent anywhere, and the text simply appears in the field. It works in Chrome and Safari; a browser that cannot do it says so instead of recording.

The service has no models of its own, so the text of the conversation goes to the OpenRouter aggregator, which passes it on to the developer of the model you chose. These are foreign companies, and in processing your text crosses the border. Who you are is not disclosed to them: neither your email nor your account identifier is in the request. Every request carries a requirement not to use the conversation for training.

Hence the usual rule for any AI chat: passwords, other people's secrets, card numbers and work documents do not belong there. The details are in the privacy policy.

Both buttons are on the Account page, and both work immediately, with no email and no waiting. "Download my data" builds a zip archive: all your chats in readable form and the same content as a machine-readable file, plus your files and images in the same folders. The archive can be built a few times an hour, no more often than that.

"Delete account" erases chats and messages, files and images, sessions on every device, your email and the password fingerprint — with no way back; a registered user is asked for the password. A guest has both buttons too, and for a guest they are the only route. Export first, delete after.

If you never signed up, this is how guest mode works, and there is little pleasant about it. A guest is recognised only by a random token in the browser: clearing the site data, a private window, a different browser or a different device all mean the conversation is not there. There is nothing to restore it with — the service has no way to know those chats are yours, a guest has no email, and writing to support will not help here. The only protection is to sign up in advance.

For a registered user the chats belong to the account: sign in with your email and password. A session left unused for half a year stops working — then an ordinary sign-in is needed.

The sign-in window has "Forgot your password?": a link is sent to the address you signed up with. It works once and expires in an hour. The service does not know your password — only its irreversible fingerprint is stored — so the old one cannot be recovered, only a new one set. Your other devices are signed out at that point.

No. A provider failure costs no credits: only what was actually generated is charged. The service first tries a backup model on its own and says so right in the chat; if none of them answered, the message that arrives states plainly that no credits were charged. The "Another take" button under an answer asks again from scratch.

Write to support@easygpt.app; the same address opens from the "More" menu under "Support". If something is broken, include a link to the chat: it shows which model answered and what was charged.