Knowledge

You are probably on Balanced already

Click the model name at the top of a chat to open the model selector. Under All, its first row is one wide card split into Eco, Balanced and Premium, for Quick questions, Everyday work and Hard problems. On a phone the three stack as rows in one box. None of them is a model. Each one lets Lexika choose the model for you.

Lexika remembers your pick on this device. Unless you picked something else, every new chat starts on Balanced, and the top of the chat shows it with Lexika under the name.

If the card is missing, select All at the top of the model selector's left column (on a phone, only its grid icon shows) and clear the search box. The card appears only there.

How the choice is made

Each message you send through a tier gets its own choice, so two answers in a row can come from different models.

Lexika chooses only from your pool, the models that pass every check:

• not in the model selector's Legacy models section
• on the workspace's Allowed models list, if the workspace owner has set one in Workspace Settings
• accepts the files on this message, with room for the message and its files
• in your plan, so its card does not show Not in your plan
• its Price is within your balance, so its card does not show Needs … stars
• its card does not show Unavailable

In a workspace on Owner Credit, the owner's plan and balance decide the pool. A low balance shrinks it, so even Premium may only reach models with a low Price. To see roughly what is in your pool, turn on Show available models with the filter button beside the model selector's search.

Lexika then reads your message and the recent conversation, along with any summary, your Chat instructions and any Brain attached to the chat. It compares them with what each model is good at and with its Price. Eco leans toward the lowest Price that does the job. Balanced weighs quality against Price. Premium goes for the most capable model, even at a higher Price or a slower answer.

To steer the choice, say plainly what the task needs. Lexika built the choice to ignore a request in your message for a particular model, so pick that model yourself.

Lexika aims to pick a model that makes pictures only when you ask it to make or edit one. Describing an image or writing an image prompt stays with a text model.

Why the chat moves to a specific model

Each answer's header names the model that wrote it, never the tier.

You may come back to a Balanced chat and find that model's name at the top instead. Match model to conversation does this. It is on by default, in Settings > Preferences under Conversation behavior. When you open a chat or reload the page, it selects the model that wrote the latest answer, so your next messages go straight to that model with no new choice.

To hand the chat back to the tier, select the tier again. The swap repeats each time you reopen the chat unless you turn the setting off, which applies to this device only. New chats still start on your own pick. The Knowledge Base article on switching models in the middle of a conversation covers what carries over.

What a message on a tier costs

A message sent through a tier carries two charges:

• The answer. Lexika bills it as the chosen model, at that model's rates.
• The routing charge. Lexika charges a small amount each time a tier makes a choice, including on a retry or an edit sent to a tier. Lexika skips it when no model passes the checks or the choice never comes back.

Lexika takes the routing charge before the answer starts, so it stays if the send then fails.

Both show in Settings > Usage. From a chat, click the star ring beside the settings gear at the bottom of the sidebar and choose View usage details. The answer's row uses the model's catalog name, which can be spelled differently from the name on the answer. The routing charge has its own row, AUTO_MODEL_ROUTING. Switch the table from Individually to Totally to add those rows up for the chosen period, Last 7 days unless you change it.

On Owner Credit, both rows appear in the owner's Usage, not yours. The Knowledge Base articles on Owner Credit and Individual Credit and on what a message costs cover the rest.

What works differently on a tier

• No Reasoning effort. The Reasoning effort control leaves the composer, and the chosen model uses its own default level.
• Files. You can attach PDFs and images only, at most 3 files of up to 20 MB each.
• No warning before you send. The composer never warns about plan or stars for a tier, so those problems show up only after you send.
• Regenerate and edit. The Regenerate list offers no tiers. Retry with same model reruns the model named on the answer, if that model is still available to you. Editing a question switches the top of the chat to the model that answered it, so pick a tier there before you send the edit if you want a fresh choice. The Knowledge Base article on Edit, Regenerate or Branch covers the rest.

If the send fails

Lexika marks your message Message failed to send and puts the reason under it. Retry is the red circular arrow beside the message.

• This model is temporarily unavailable. Please try another model. Lexika could not make the choice, or every model that passed the other checks shows Unavailable. Nothing falls back to a default model. Retry, or pick a model.
• Model not found. Please check the model ID. No model passed the checks, for example because your balance (the owner's, on Owner Credit) is below every Price, or the Allowed models list leaves none. Lexika may also have found no suitable model. Pick a model yourself, and the composer shows what stands in the way. In an Owner Credit workspace you don't own, it says the workspace's plan doesn't cover the model, even when the owner is only short of stars. The model's card in the model selector shows Needs … stars when the owner's balance is the cause.
• You don't have enough stars. Lexika holds stars for reading the conversation plus the model's Price, and in a long chat that can exceed your balance. The routing charge is already taken. Add stars, or start a new chat.
• NO_ATTACHMENT_COMPATIBLE_MODEL_FOR_AUTO. No model Lexika could choose accepts these files. Remove them, or pick a model that takes them.
• NO_IMAGE_COMPATIBLE_MODEL_FOR_AUTO. You asked for a picture, and no model in your pool makes pictures. Pick one yourself.
• MODEL_CONTEXT_WINDOW_EXCEEDED. Your message and its files are too large for every model Lexika could choose. Shorten the message or attach fewer files.

When to pick a model yourself

Pick a model when one model should stay with a long piece of work, or when you need Reasoning effort, other file types or a warning before you send. Pick one too when your balance is low and you would rather choose than take whatever is still covered.