Auto mode: a fast model for simple things, a strong one for hard ones

1How Auto chooses a model

The decision is made on the server from the words of your message and its attachments, without a separate AI call. So Auto adds nothing to the price and doesn't slow the reply down.

The strong model is switched on by these signals:

• a file or an image is attached to the message;
• the topic is video, documents and tables, a storyboard, creating a persona, souls and agents, mailings, your account, payment or support, or you explicitly ask to draw a picture (the assistant recognizes the topic by keywords);
• you ask for a schedule or a repeat (“remind me”, “every day…”);
• the text has code or programming words (function, error, Python, SQL and similar);
• you ask it to think or compare (“think”, “compare”, “analyze”, “pros and cons”);
• a long conversation (more than about 24,000 characters of history) or a very long message (more than 6,000 characters);
• Research mode is on (a report from many sources).

Words about pictures and schedules raise the model only as an explicit request. For example, “Translate: the meeting is tomorrow at 10 am” stays on the fast model: the word “tomorrow” alone is not a signal.

Once a conversation has become “strong”, it stays strong for 6 hours: without a new signal the model is not raised, and in the middle of a hard conversation it is not quietly downgraded. If a model is locked by your plan, Auto takes the next one in its list.

2The line under a reply: who answered

Under every reply in Auto mode there is a small grey line. For an ordinary message it reads “Auto · fast model · Gemini 3.8 Flash”. For a hard one: “Auto · strong model (code) · Claude Sonnet 5.5”. The reason is in brackets: “file”, “tools”, “code”, “hard task”, “long text” or “continuing the thread”.

Hover over the line and a hint appears: “Auto picked a model for this message. To turn it off, choose a model in the list.” The Continue button (when a reply was cut off) continues the reply with the same model that started it.

3How to pick a specific model or go back to Auto

1. Open the mode menu

   In a new chat, click the button under the input field on the right. It says “Auto”, “Fast” or “Thinking”: that is the thinking-depth mode, not the choice of model.

2. Click the “Model” row

   A list of models by developer opens. At the very top there is a separate Auto row with the caption “Picks for you: a fast model for everyday questions, a strong one for files, tools and hard tasks” and a price range.

3. Choose Auto or any model

   The choice is remembered: for this chat and for new chats. With a model you chose by hand, Auto stays out of the way and the “Auto · …” line doesn't appear under replies.

If you picked a model by hand earlier, it stays selected: Auto is the default only where no choice was made yet. Auto works in the web version; the Gibrid mobile apps still answer with their previous default model. To compare several models on one question, see the page about comparing models.

4What it costs: the “≈ N G” chip by the send button

To the left of the send button in the input box there is a small chip with the price of the next message, for example “≈ 0.5–1.6 G”. G is the same as the ⚡ token (1 ⚡ ≈ $0.025). The price is calculated in advance by the same formula that is later used to charge your balance:

• the assistant's built-in part of the request (its instructions and tool descriptions): about 12,000 tokens;
• plus the history of this chat (about 1 token per 3.2 characters) and attachments (1,500 tokens per attachment);
• plus a typical reply of this model; the price comes from the model's price table with your plan's markup;
• a per-model correction: once an hour the server compares the estimate with actual charges over 14 days.

For Auto the chip shows a range from the fast to the strong model. The exact amount is only known after the reply: it depends on which tier Auto picks and on the length of the reply. For a model you chose by hand the chip shows a single number. The same price (“≈ N G in this chat”) is shown on every row of the model list, so you can see an expensive model before choosing it.

5A reply was cut off: the “Continue” button

A reply used to be cut off mid-word because of a length ceiling. Now the ceiling comes from the model's capabilities but is no more than 16,000 tokens (up to 32,000 with deep thinking). If a reply still hits the ceiling, a line appears under it: “The reply was cut off by length · Continue”.

1. Click “Continue”

   The assistant finishes the reply from where it stopped, without repeating itself and without a greeting.

2. Keep reading in the same bubble

   The continuation is added to the same reply. The service phrase “continue” does not go into the chat history.

The same model that wrote the beginning writes the continuation. The charge is fair: only the continuation's tokens are charged, and in the “cost of the reply” they add up with the first part. The button appears only on the last reply of the conversation.

6A long conversation: the start is condensed into a summary

The model doesn't see the whole conversation, only the latest messages that fit its window (the web chat sends up to 30 messages). Before, everything earlier simply disappeared, and the assistant answered “as I said at the start…” off the mark. Now whatever didn't fit is condensed by the inexpensive Gemini Flash model into a short summary (no more than 1,200 characters) and passed to the assistant along with your question.

Under the reply a line appears: “The start of the conversation (24 msg.) was condensed into a summary · Show”. Click Show to read the summary itself and Hide to collapse it.

• the summary is saved and is recalculated not on every message but when about 12 new “dropped” messages have piled up;
• the cost of condensing is included in the cost of your reply; there is no separate line in your balance;
• in a temporary chat (covered on its own page) no condensing is done, because the history is not saved;
• there is no “new chat with a summary” offer: when a message gets expensive (10 G and above), the price chip itself suggests starting a new chat.

7Feature hints under a reply

Sometimes a single short hint appears under the last reply. Each kind is shown no more than once a day, and you can close it with the cross:

In a temporary chat the hints are not shown. More about public links on the page about sharing a chat.

8What Auto does not do

• It doesn't read minds. The choice goes by words and attachments. If a hard question is written very briefly and without “signal” words, it can go to the fast model. Then clarify the task or pick a strong model by hand.
• There is no “always strong” button. If you want the same level every time, choose a specific model in the list.
• It doesn't work in the mobile apps. For now Auto is in the web version gibrid.ai/chat.
• It doesn't guarantee an exact price. The chip is an estimate; the real amount is visible after the reply: in the “⋯” (More) menu under the reply there is a “Cost ≈ N G” line.

Related