Knowledge

You picked a different model halfway through

You changed models partway through a chat, and something is off. The new model claims it wrote an earlier answer, a PDF from the start seems forgotten, or the Brain stopped answering. Or the model changed by itself when you reopened the chat.

To switch, select the model name at the top of the chat to open the model selector, or type /model at the start of the message box and choose Change model. Your next message goes to the model you pick.

If you use Eco, Balanced or Premium, the tiers that pick a model for you, Lexika may hand each message to a different model. Most of what follows can then happen without you switching. The Knowledge Base article on Eco, Balanced and Premium explains how they choose.

What the new model reads

The new model reads the same conversation the previous one did. That means the latest compaction summary, if there is one, then every question asked of the AI after it, by you or a teammate, and every answer. Comments are left out.

• The new model takes earlier answers as its own. They arrive as the assistant's turns with no model name attached. The new model cannot tell that another model wrote them, and it may say it wrote them itself.
• The new model reads answers, not the work behind them. Lexika never saves a model's reasoning, and it does not pass on the results of earlier web searches or Brain lookups. The new model sees only what the answers said about them. Pictures that earlier answers made do reach it, as files. A model that can use tools also rereads what connected apps returned in the last four answers.

When the new model can read less

Models differ in how much they can read at once, their context window. Lexika does not show it for any model. Before every message, Lexika fits the conversation to the model that will answer. When it does not fit, Lexika leaves things out in this order:

1. Files and pictures in earlier messages, oldest first.
2. What connected apps returned in earlier answers. The answers' text stays.
3. Whole earlier exchanges, oldest first.
4. The compaction summary.

Nothing on screen says this happened, and your transcript stays complete. If your new message and its files still do not fit, the message shows Message failed to send with MODEL_CONTEXT_WINDOW_EXCEEDED under it. Remove some files, shorten the message or try another model.

Files after a switch

An earlier picture or PDF reaches the new model only if that model accepts the file's type and size and has room for it. Otherwise the model gets a short note naming the file. Whatever the model, Lexika carries only the 10 newest files in a conversation, your new message's files included, and leaves older ones out without a note.

Lexika checks files waiting in the message box against the new model the moment you switch or @mention one, text files included. It removes any the model cannot take and shows a short notice naming the model. The Knowledge Base article on files in a chat covers the limits.

Models that cannot use tools

Tools are the actions a model takes besides writing, such as searching the web, reading the Brain or using a connected app. Models that answer with a picture cannot use any, and neither can Perplexity's models. The model selector does not say which other models cannot. A model without tools:

• does not use Lexika's web search, so no Searched the web row appears. Perplexity's models search the web on their own, so their answers can still draw on it.
• cannot read the Brain or save to it
• cannot use connected apps. An app you chose for the conversation keeps its chip in the message box, but this model ignores it.
• does not react to your messages or make pictures partway through an answer

In a chat with a Brain, Lexika asks such a model, if it answers in text, to send you to a model with tools when you ask it to remember or recall something.

A chat has no web search switch. A model that can use tools decides for itself when to search, and you can ask it to. If you ask and no Searched the web row appears, pick another model and ask again.

Which model answers your next message

The name at the top of the chat shows where your next message goes, unless you mention a model. Lexika decides in this order:

1. An @mention. Type @ in the message box and pick a model under Route this message to a model. That one message goes to it, and the name at the top does not change. A message can carry only one @mention.
2. The question you are editing. While you edit a sent question, the name at the top shows the model that answered it. The Knowledge Base article on Edit, Regenerate or Branch covers the rest.
3. The model of the latest answer, when Match model to conversation is on, until you pick a model yourself in that conversation.
4. Your last pick in the model selector. Only this browser or app keeps it, so your phone, your laptop and a second browser can show different models. New chats start with it.
5. Balanced, if you have never picked a model or your last pick is no longer offered.

If the mode switch in the message box reads Chatting instead of Asking AI, a message reaches a model only when you mention one. Your pick steers only your own messages. A teammate in the same conversation sends to whatever their own screen shows.

Match model to conversation

The switch is in Settings > Preferences under Conversation behavior. It starts on, and each browser or app keeps its own setting. With it on, opening or reloading a conversation selects the model that wrote its latest answer. That answer may be a teammate's, or one you got through an @mention.

If Eco, Balanced or Premium chose that model, you land on the model, not on the tier. To go back, open the model selector and pick the tier at the top of its first tab, All, which shows only an icon on a phone. The next time you open the conversation, you land on the tier's latest choice again. Turn the switch off to have every conversation open on your own last pick, a tier included.

If you cannot use the selected model, a notice above the message box says why, with an Upgrade plan or Get stars button when one would help. In an Owner Credit workspace you do not own, it says the workspace's plan does not cover the model, even when the owner is only short of stars.

Telling which model wrote an answer

Every answer starts with a header showing the company's logo and the model's name, with the company and the time beneath. Switching never changes the name on an earlier answer. For a model Lexika no longer lists, the header shows the stored name with a plain logo and only the time.

Switching without surprises

• Restate what matters after a switch. If a long conversation depends on an early decision, repeat it. Attach an early file again if the new model takes that kind of file, or paste the part you need as text.
• Expect the Reasoning effort control to move. It shows the nearest level the new model offers, or disappears when that model offers no choice. Your own level returns when you switch back, unless you changed it in between.
• Count the cost of a late switch. Lexika charges each answer for everything the model reads, at that model's rate, so a pricier model makes every later message in a long conversation cost more. Compare models by the Price in the model selector. The Knowledge Base article on what a message costs covers the charge.