Agent chat
Models
The model menu, what a turn costs and why, and how to choose.
Every chat runs on one AI model at a time. The model does the reading, reasoning and writing. Agent Hero gives it the tools, documents and approvals.
Switch the model
- Click the turn-settings button above the message box. It shows a shield and the model's mark.
- Under Select model, pick a model.
- Send your next message. It runs on the new model.
You can only switch between turns. The list is greyed out while the agent works.
New agents start on Claude Opus 5.5, unless you picked another model before. Your pick becomes your default for agents you create afterwards. Existing agents keep their model.
A badge such as 20% off marks a discounted model. Every turn on it is discounted.
What changes when you switch
Stronger models plan better and slip less on messy documents. Lighter models answer faster and cost far less. The agent's documents, skills, connections and approval settings stay the same.
Tokens
Models read and write in tokens, chunks of a few characters. A page of text is roughly 500 tokens.
Input is what the model reads: your message, the conversation so far, documents and tool results. Output is what it writes. Each model charges its own price per million tokens.
What a turn costs
A turn is your message, everything the agent reads and does to answer it, and its reply. It costs each kind of token times the model's rate, added up, and is paid from your workspace's credits.
The reply's ⋯ menu shows the turn's Cost, its Input, Output and Cached tokens, and the model that ran it. Cost is left out when it is $0. Cached only appears when some context was reused at a lower rate.
The $ left chip by the message box shows what you can still spend. That is the workspace balance, or less if an admin set you a spend limit.
Why some turns cost more
- Long chats, because each turn re-reads the conversation
- Big documents read in full
- Multi-step work with many tool calls
- Stronger models, which cost several times more
- Long answers, since output costs more than input
Very long chats summarise their older parts to keep working. A new task for a new topic is still cheaper.
To keep costs down, start a new task for each new topic. Upload or @-mention a document instead of pasting its text. Use a lighter model for routine extraction and formatting.
Compare prices
Click See prices in the menu. It opens Settings → Models, with each model's Intelligence Index and its input, output and cache prices per million tokens. Click a column to sort by it.
For the full price list, credits and plans, see Billing.
A simple way to choose
Stay on the default until you have a reason. Claude Opus 5.5 is one of the strongest models in the menu, and priced to match.
| Situation | Move |
|---|---|
| High-volume, mechanical work | A light model: Claude Haiku 4.5, Gemini 3.8 Flash, GPT-6 Luna, DeepSeek V4.1 Flash |
| Long sessions you'll iterate heavily | A mid-tier model: Claude Sonnet 5.5, GPT-6.1 Sol |
| Hard reasoning across documents | Stay on the default |
| Not sure | Default, then judge the output |
The menu also lists Gemini 3.1 Pro (Preview), Grok 4.7, MiMo V2.6 Pro and others. Try one on a task you can check first.
If a model runs out of capacity, the chat asks you to pick another model and send again. The work saved so far is kept.
When a model retires, its chats move to its replacement. Skills pinned to it move to a newer model of the same family, or else to the default.
A skill can pin a model. Pin routine skills to a light model and the policy enforces itself.
Muse Spark 1.3 Contributor is cheap because Meta's contributor terms let Meta use what you send it to improve its products. The app doesn't warn you, so don't use it for client or other confidential work.