Frequently asked questions

BAIchat is an app for roleplay chats with AI characters, 18+. It runs in the browser at baichat.app and as a Telegram mini app.

• Pick a character or a scenario in the Library and start a chat.
• Not sure how to begin? Tap "Intro?" and the AI writes the opening scene.
• As the story goes on, the character remembers what matters, your relationship changes, and events build up in the chronology.
• Turn any scene into an image, animate it into a short video, and publish your best work in the Gallery, which runs a contest every week.
• Create your own characters and scenarios in the Workshop.

You get 600,000 tokens when you sign up, no card required. Messages and images use tokens; when they run out, you can top up. There is no free mode and no daily bonus; the "free" entry in the model list is left over from old versions and also uses tokens.

Thousands of characters and scenarios: official ones curated by the team and ones made by the community. Search by name or keyword and filter by popularity, newness, gender and creator. The "Franchises" tab groups characters by anime, game and film, and you can vote for new franchises.

Settings has a description of your appearance: characters take it into account in conversation and in images that include you. You can make several appearance presets and switch between them in different chats. You can also put your name in a preset, and characters will call you by it.

No, a photo can't be used as your appearance or as a reference for image generation. But in the preset editor you can upload a picture and the AI writes a text description of it: appearance, outfit, or both. The image is not stored anywhere: it is processed and discarded right away, and only the text, which you can edit, stays in the preset. The analysis costs a few tokens.

Yes, in group chats: up to six characters in one scene. They talk not only to you but to each other. See "Group chats" for details.

In any browser at baichat.app, the main version, and in the Telegram mini app. If you link your accounts, chats and balance are shared.

• Actions, gestures and the setting go in asterisks: walks to the window and looks at the sunset.
• Speech is plain text.
• Thoughts also go in asterisks, for example wonders whether to trust him. Characters react to what you visibly do, but they don't hear your thoughts.

Characters mirror your style: detailed messages get detailed replies. Instead of "Hi", write something like: smiles as I greet her at the door Hey! It's been a while. hugs her How was your day?

• Describe the scene: where you are, what's around, who is holding what.
• Give context about yourself and the situation: who you are, how much time has passed, what happened.
• Act and decide yourself: a character can move the story along more easily when something is happening.

Tap "Intro?" next to the message field: the AI writes an opening scene based on the character, your appearance and the selected scenario. Or describe the start yourself to set the place, time and roles right away.

• Send: with text, sends your message; with an empty field, opens "Generate reply", where the AI suggests your next line from the conversation. You can edit it before sending (this uses tokens).
• Continue: the character carries on the story without a message from you.
• Regenerate: a different version of the character's last reply.

Yes. You can edit your own message and regenerate the character's reply. You can roll the chat back to any message or branch a new chat from it, keeping the old one. You can keep several separate chats with the same character and export a chat.

The easiest way is through action in the scene itself: characters react to what happens. If the character got something wrong, correct it briefly in your message, for example (he wasn't wearing a jacket). Don't turn your messages into long instructions for the AI: that usually makes replies worse.

A message can be up to 4,000 characters, but it's better not to get close: in a very long message the model has a harder time seeing what matters, and the reply gets worse. Backstory and important world details are better placed in a scenario.

A scenario sets the stage: the setting, your role and your relationship with the character. You pick one in the chat menu; your own scenarios are made in the Workshop.

Characters marked "+universe" can draw on detailed knowledge of their world and other characters from their franchise. Turn it on in the chat settings; you can enable a universe in any chat, even with a character from another franchise. Replies take 5–15 seconds longer.

There is a copy button next to each message. In the Telegram mini app it sometimes doesn't work because of Telegram client restrictions on some devices; in the web version it always works.

Character replies and image and video generation. Besides the reply itself, the memory system occasionally uses a few tokens: compressing older messages into the chronology and picking relevant memories. This happens once every few dozen messages and is sometimes charged one or two messages later.

To see the model and cost under each reply, turn that on in Settings.

Every reply is generated by a large language model running on expensive hardware with limited capacity. Each message costs us real money, so chatting can't be free and unlimited. To make tokens last longer, pick a cheaper model or the "Economy" memory mode.

Usually 1,000 to 8,000 tokens, depending on the model. Most of the cost is not the reply itself but what the model reads before replying: the recent conversation, memory, the character description and the scenario. That's why the reply's length barely affects the price, while the cost varies from message to message. Regenerating is cheaper than a normal message, and the "Economy" memory mode lowers the cost.

Roughly 10,000 to 35,000 tokens, depending on the model and quality. You need at least 10,000 tokens on your balance to start a generation.

No, tokens never expire.

Sometimes: promo codes, the weekly lottery and Gallery contest prizes. The referral program is currently off.

Enter the code on the top-up page before paying: the bonus is added with your purchase. Each code can be used once. The promo bonus stacks with the package discount; the token amount shown on a package is already final.

Tokens usually arrive within a few minutes, sometimes up to half an hour. If it takes longer, contact support with your account ID (shown in Settings) and the time of payment.

The price is tied to the exchange rate of the payment systems and may be adjusted along with it.

The draw takes place every Friday. You earn tickets for activity: sending, regenerating and continuing messages, and generating images. How many tickets you need to enter and what the prizes are is shown on the lottery page (tap your ticket count in Settings). After each draw, everyone's tickets reset.

Every Sunday the authors of the week's best publications, by community vote, receive tokens. Winning works stay in the "Winners" filter.

For most people the default "Auto" mode is best: a mix of GLM 5.2 and Gemma 4. GLM writes better, while Gemma is cheaper and keeps the mix steady. If a chat starts acting up, you can switch to GLM 5.2 alone for a few messages and then go back to the mix.

When the main provider is overloaded or temporarily not responding, the request goes to a backup model so you don't get an error. This happens more often at peak hours and usually passes quickly. You can see which model replied under the message.

The model is temporarily unavailable at the provider. Regenerate the reply or pick another model with the "Model" button.

The model menu has a reply length setting: "Auto", "Short" and "Long". "Auto" adapts to the scene. Even on "Long", the model won't stretch a reply if there's little going on in the scene: move the story forward or give it more detail.

Sometimes a model gets stuck like this. Regenerating, rolling back one message, or switching models helps.

This is a quirk of the models themselves. If a word gets in the way, regenerate the reply.

Strong repetition penalties break grammar, especially in Russian, so repetition can't be removed completely. It's lowest when several models are mixed, as in "Auto" mode.

A single strange reply is usually a glitch: regenerate or switch models. If refusals keep happening, check the character's description: sometimes it's written in a way that leaves the model unsure what's expected in such a scene.

New models are tested regularly, but only those clearly better than the current ones in quality and price are added.

No. Characters rely only on what's in the app: the character description, the scenario, your conversation, memory, and universe knowledge if it's on. They don't know current news, the weather or what's behind a link, and if you ask, they may guess.

Each character has familiarity, affection and trust bars (from −100 to 100), plus a mood and a relationship status. They update as the chat goes on and shape how the character talks to you.

Relationships have to be earned, and characters judge you from their own point of view: a villain won't warm up to someone hunting them. 100 is meant to be nearly unreachable; 90 is already a lot. Bad actions and conflicts lower the bars.

Difficulty (easy, normal, hard) sets how easily relationships grow and how sharply they drop. Choose it in the model menu in a chat. The difference shows over time, not in a single message. In a group chat, difficulty is shared by all characters.

Yes: in a new chat, before the first message, you can set any relationship values and status. After the first message, only the chat itself changes them.

The model sees the latest messages in full. Older ones are periodically compressed into the chronology, a list of key events, and each reply pulls in the memories relevant to the current scene.

The chronology is a compressed summary, so small details fade over time. If the character forgot something important, just remind them in a message. Wrong entries can be deleted from the chronology, and anything that must always be remembered fits well in a scenario.

You can delete chronology entries, but not edit them.

Up to six. Usually one or two characters reply to each of your messages, sometimes three; the scene director decides who.

Tap the double-arrow button: the characters carry on the scene. You can pick who speaks next or leave it on "auto".

Remove them from the members and add them back later: their memory and relationship are kept. You can also just send the character somewhere in the story: they reply and leave the scene, and won't know what happens without them.

Yes, whisper it in an action, for example whispers in Gale's ear …. The other characters won't hear what was whispered.

You can regenerate the last reply or the whole turn. A reply in the middle of a turn can't be regenerated, because the characters after it reply with it in mind.

Yes, you can create a group chat from an existing chat without starting the story over.

Every chat has an "Image" button (at the bottom or on the side, depending on your device). The image is based on the current scene; you can choose the model and style. A finished image can be animated into a short video.

Image generation is restricted by the rules of payment systems and ad platforms, and the automatic check sometimes errs on harmless scenes. Try another model: Seedream often works well for horror and dark scenes. If an image is blocked after it has been generated, the generation is still charged.

Image models sometimes drift from the description, especially in details (ears, tails, hair color). Check your appearance preset or create a new one; for a character, an avatar cropped to the face helps.

Yes, but results aren't guaranteed: image models don't always handle several characters in one scene.

Open the image and tap "Publish". Only images with public characters can be published. Moderation is automatic; the rejection reason "Other" means the image doesn't fit the Gallery's format or rules.

If a video made from an image is published in the Gallery, the image itself won't pass moderation. You can publish one or the other: the image or the video made from it.

In the gallery on the chats page. Images are kept even if you delete the chat itself.

In the Workshop. "Smart" mode is for well-known characters: the AI gathers the description itself. Manual mode is for your own original characters: describe their personality and looks, and the AI helps fill in the card. The cost is shown before you create.

Usually a few days. While a character is in moderation, you can already use it yourself, including in group chats; it appears in the public Library once approved. You get a notification with the result. Editing a character sends it back to moderation.

The avatar must be cropped to the face, like official characters: in a wide shot the face is too small, and image generation reads it worse.

"My characters" shows a character code. You can share a private character with that code; your friend keeps it even if you change the code later.

Suggest it in the Library: the community decides by voting. Franchises are added gradually, and not every character makes it in: a character needs enough official information.

Scenarios are usually 5,000–7,000 characters. It's best not to go beyond 15,000–17,000: the longer the scenario, the more expensive and less stable the chat.

Create a code in the settings of one version and enter it in the settings of the other. After linking you have one shared account with the same chats and balance.

Contact support with your account ID from Settings, and the link will be undone manually.

A login lasts 14 days; after that you need to sign in again.

Open the web version in your browser: baichat.app. It is more stable, especially with a poor connection; turning on or switching a VPN sometimes helps.

No, several accounts per person are not allowed and can lead to a ban. To share characters, use the character code.

There is no separate app. You can add the site to your phone's home screen, and it opens like an app.

No, chats are private and never published. They are stored on the server so memory works, and they are deleted along with the chat.

Most often it's a short restart or maintenance, usually 5–20 minutes. Wait and refresh the page. We announce major outages in the outages topic of our Telegram group.

On the connection screen, open the technical details, copy the text and send it to support: sometimes your region needs a dedicated proxy. Until then, a VPN or the web version helps.

It's loading the latest version of the interface: Telegram doesn't keep it between launches. The reload is harmless.

It appears when a regenerated reply came from a different model. Nothing is wrong.

Post in the "Bugs" topic of our Telegram group: which model, which chat (one-on-one or group), what happened and, if possible, a screenshot. If it's about your account or a payment, include your ID from Settings.