AI Chat

The chat gives you access to text neural networks: ask, explain, translate, write code, go through a document or a picture. It differs from the generator fundamentally: the chat is text-only, there are no "Images" or "Video" tabs here, even though the screen layout looks similar.

Three parts: the list of conversations on the left, the conversation in the middle, the settings on the right.

Here is how it looks on a phone: the side menu opened straight from the chat, every section of the site inside it, and at the very bottom, below the rule, the globe with the current language.

The right panel: four blocks

Model · Ready-made prompts · System prompt · Additional. The last block does not always appear — it is assembled from what the selected model can do.

Choosing a model

A list of models with search and tags, as in the generator. Text models are priced per thousand tokens, and the price is on the card.

Ready-made prompts: roles for the assistant

These are not sample questions but system prompt templates — ready roles for a data analyst, a product manager, a technical writer, a contract editor, a scriptwriter, a translator, a senior developer and more. Their names and categories stay in Russian on the English locale, so go by the card's description. Roles have search, categories and bookmarks: a marked role rises in the list, and the star to the left of the categories leaves only the marked ones — the same trick as with styles in the generator.

Example. The "translator" role sets the behaviour once — after that you just drop text into the chat, with no "please translate" in front of every chunk.

Tip. A role is a starting point, not a restriction: pick a ready one and add your own lines to the system prompt, they don't conflict.

The list of roles is loaded from the server, so it has a loading state and an error state with a "Retry" button. The list may grow, and its height is limited — if it looks like there are only three roles, scroll.

The system prompt

The field reads "Set the assistant's behavior…", and its tooltip spells it out: "Instructions that define the model's behavior at the start of each conversation". Next to the block heading sits its own cost counter and a token estimate — something like "1 · 23 tok." — so you can see what the instruction itself costs.

The eraser on the right of the field appears only when there is something in it, and it also clears the selected role — text and choice at once. There is no separate "remove role" button, which is not obvious.

Additional: answer parameters

• "Max output tokens" — a slider whose bounds are set by the model.
• "Response verbosity" — three buttons: Concise · Balanced · Detailed; the middle one is picked by default.
• «Рассуждение» (reasoning) — a toggle on models that can do it, and one of the labels left untranslated. Turning it on reveals the reasoning effort steps.
• «Инструменты» (tools) — a toggle on models that support tools, on by default; also untranslated.

If a model already describes its reasoning in a scheme of its own, no separate toggle appears: the service does not show two identical controls.

Reasoning helps on logic problems and gets in the way on simple ones, where it only burns tokens. For "rewrite this email more politely" it is better off.

Tools: when the answer is drawn, not written

With tools on, the model can answer not with text but with a widget: a chart or a table, a plan with step progress, source cards, a list of options to choose from, an order summary, a carousel of cards, an image, a video or audio player, points on a map, terminal output.

Example. "Compare three plans by price and limits" with tools on gives a real table rather than text with asterisks. "Plot sales by month" gives a real chart.

What else an answer can render

An answer is not plain text. It supports GitHub-style markup (tables, checklists, strikethrough), mathematical formulas instead of raw LaTeX, and code with highlighting and a copy button.

Buttons around the input field

To the left of the field — a paperclip and a web search icon. To the right — a microphone and the Send arrow, with the message cost counter beside it.

Web search

The globe icon makes the model go to the internet before answering. The tooltip spells it out: "Search the web before answering. Applies to one message, +N coins per request" — the surcharge is written right there.

An answer made with search arrives with links to its sources — they are visible right in the thread, so the numbers can be checked.

Some models search the internet by themselves — they have no toggle and don't need one.

Attachments

You can attach with the paperclip or by dragging files into the window — while dragging, a "drop the files to attach" hint appears. The service lists the formats itself, in the paperclip's tooltip: images PNG, JPG, WEBP, GIF; PDF documents; text and code TXT, MD, CSV, JSON, YAML, LOG and others; audio MP3, WAV, OGG, FLAC, M4A; video MP4, MOV, WEBM.

Every file has its own status — queued, uploading, uploaded, error, cancelled; an upload can be retried, cancelled, or the file removed. Removing works only for what has not gone into the chat yet: a sent attachment disappears together with its message.

An attachment can be refused for three different reasons, and each is cured differently:

1. too many files in one message — the limit is substituted into the message; when dragging a batch, only the first ones are added;
2. the format is not supported at all — the message lists what is allowed;
3. the model cannot read such files — the format is fine, but the selected model does not work with it. Only changing the model helps here.

Attachments are estimated in advance and cost coins: if the balance is short, the service says outright what the files were estimated at and offers to remove some or top up.

What a message costs

The mechanics are the same as for image generation: before sending, coins are reserved with a margin — based on your text, the system prompt and the maximum answer length. After the answer, the actual usage is charged and the difference returns. Attachments add a separate sum, and reasoning and web search make a request more expensive.

A live answer and what is under it

The answer types itself out in front of you. Under it — the token count and the generation time; an answer made with search carries a note saying so.

A message has seven states: sending, request accepted, generating an answer, recovering the answer, answer ready, generation error, cancelled.

Actions on a message

Copy · Download as Markdown · Regenerate · Send again · Retry · Edit.

Three cases where the buttons silently do nothing:

• "Retry" works only for a failed or cancelled answer. On an ordinary finished answer it will not fire;
• any action is blocked while an answer is being generated in this conversation: one conversation, one stream — a second answer will not start alongside;
• if the message had files and they are no longer available, retry and edit will not work — an attachment lives for one message. The way out: send a new message and attach the files again.

"Edit and send again" cuts the conversation from that message onwards — everything that came after disappears. This is a common loss: someone edits an old question and loses the whole branch below it. Editing applies to your own message only: an assistant's answer is changed with "Regenerate", not by editing.

An answer can be stopped mid-sentence, and nothing is lost. The part already written is kept, the message is marked "Cancelled", your question stays in the thread, and "Retry" appears next to it. Pressing stop is not scary.

An empty message cannot be sent, but a message consisting only of attachments, with no text, can.

Conversations

On the left — the list: search, grouping by Today / Yesterday / N days ago, renaming, deleting and "New chat". An unnamed conversation is called "New conversation".

There is also "Clear conversations" — deleting them all at once.

If the chat answers with an error

Message	What it means and what to do
"The current message does not fit into the selected model's context"	the conversation or attachment is longer than the model can read — start a new chat or take a model with a bigger context
"The selected model cannot work with images — choose another model or remove the picture"	the message names both ways out itself
"This file format is not supported"	the list of allowed formats is substituted into the message
"No more than N files can be attached to one message"	the limit is supplied by the server
"One or more chat attachments were not found, already used or unavailable"	an attachment lives for one message; attach the file again
"The attachment is too large to send to the chat"	the file will not go through whole — shrink it or send a fragment
"The chat history has changed. Refresh the chat and try again"	the conversation changed meanwhile — for example, it is open in a second tab
"Session expired. Sign in again"	the sign-in ran out of time; the thread on the server is intact
"The connection dropped before the answer finished"	the link broke mid-sentence — press "Retry"
"Parallel answers limit"	you started answers in several conversations at once; wait for the current ones
"Chat models are unavailable for now"	the model list did not load; the panel has a button to load it again
"The request is still processing" / "The request was already accepted" / "The request is already complete"	pressing again broke nothing and charged nothing twice

Part of the chat interface is not translated

On the English version these stay in Russian: the "Готовые промпты" (ready-made prompts) and "Системный промпт" (system prompt) headings, the "Рассуждение" and "Инструменты" toggles, the names and categories of roles, the welcome line, and the "Повторить загрузку доступности моделей" button that reloads the model list. Everything else — the model panel, the input field, the "Additional" block — is in English.

Go by the icons, by the position of a control and by this guide.

Where to go next

• Send your own picture to the chat with one button — Generator.
• Understand what makes up the price of an answer — Coins and limits.
• Pick a model by description and tags — Model catalogs.