Models
Decide which model does which size of job, and what chat search and image generation run on.
Feature Updated 20 Nov 2026What the Models page is
A model is the AI that writes Panda's answers. Some models are strong and slower, some are quick and light. A plan for a new feature and a one line question do not need the same one. The Models page in Settings decides which model each kind of work runs on, so you do not have to pick one for every message.
Where to find it
Open Settings and pick Models. From top to bottom the page holds Connections, Default, Usage Caps, Usage Limit Failover, Behavior, Chat search and Image generation. Where your models come from is covered in Connections.
Model tiers
Instead of one model for everything, each connection has five tiers. You set them on the connection's Tiers tab.
| Tier | What runs on it |
|---|---|
| High | Deep work: planning, review, hard problems. Agents bound to High use it. |
| Medium | Everyday work. New chats start here, and the composer changes it per chat. |
| Low | Quick, simple tasks. |
| Utility | Background jobs such as chat titles and commit messages. Nothing picks it. It runs on its own. |
| Vision | Reading images. When a chat's model cannot see, an attached image is read here for that turn only. |
Agents, automations and chats on another connection use that connection's tiers, which is why a connection you are not using still has tiers worth setting.
Auto
On PandaOS Cloud, pick Auto in the composer's model picker. PandaOS then chooses the model for each message from what the message needs, for example one that can read an image you attached. The other connections do not offer Auto: there you pick the model, or leave the one the Medium tier sets.
Which model a new chat uses
Under Default, the Connection row sets where new chats, automations, agents and background jobs run. Model for new chats sets the model a new chat starts with. Two switches under them can override that for chats. One is Reuse the model and connection from my last chat, the other Reuse the model and connection from my last chat in that project.
Usage Caps
Under Usage Caps, Enable usage caps keeps part of your subscription limits free for use outside PandaOS. It also caps what each paid connection spends per day, week or month. It is off until you turn it on. Three rows appear once it is on:
| Row | What it does |
|---|---|
| Session cap | How much of a subscription's session limit PandaOS may use. |
| Weekly cap | How much of a subscription's weekly limit PandaOS may use. |
| Notify at 80% of a cap | Tells you once per window, before a cap is reached. |
With a paid connection set up, a Spending limits box follows, with one row per connection. Each row takes a monthly limit, and its menu adds a weekly or daily one. Empty means no limit.
When a cap is reached, chats move to another connection if usage limit failover is on. Otherwise they pause.
Agent model switching
Under Behavior, Agent model switching lets an agent run on its own model, connection or reasoning effort while it is working. The chat returns to your setup afterwards. How to give an agent its own model is on Agents.
When the feature is available to you, Behavior also holds Enhanced plans. It runs bigger work in phases, such as plan, build and ship, with a check-in between each.
Which models you can pick
Which models the picker offers depends on the connection. PandaOS Cloud serves a catalog we host. Every other connection serves what your own account or computer offers. On PandaOS Cloud every model uses your plan usage; on the other connections your provider bills you at its own rates.
PandaOS Cloud
PandaOS Cloud serves a catalog we host. The app reads the list from the service, so the models in your picker are the current ones and this page does not repeat them. Where one model is available from more than one provider, the picker shows it as a single entry.
Chat search embeddings and image generation run on their own models, covered under Chat search and image generation.
Claude
The Claude connections share one catalog of Claude models. Reasoning effort is per model, not universal. The levels the composer offers depend on the model you picked, and some models have no effort levels at all.
Superseded generations are hidden from the composer until you opt back in from the connection's Models tab. Hiding is conditional: a model is only hidden while its successor is in the same list.
OpenAI Codex
Codex ships no fixed list. PandaOS reads the catalog from the Codex CLI on the account you signed in with. The models you see are the ones your ChatGPT plan actually entitles you to. Earlier generations are hidden by default once a newer one appears in the same list.
Your API keys, Local, and OpenCode Go
These three have no catalog of ours to list, because the catalog is yours.
| Connection | Where the list comes from |
|---|---|
| Your API Keys | The providers you add and the models you register on each. Because one list can mix keys, the picker groups models under the provider paying for them. |
| Local | Whatever is loaded in Ollama, LM Studio, KoboldCpp, llama.cpp, or your own endpoint. Grouped by runner, since the same model name can exist on two of them. |
| OpenCode Go | The models your opencode.ai subscription serves. |
Chat search and image generation
The bottom of the Models page sets two jobs that are not chat models. Changing your connection never changes them.
| Setting | What it does | Where it runs |
|---|---|---|
| Chat search | Finds past chats by meaning, not only by exact words. The Atlas row picks what indexes your knowledge pages; it follows chat search unless you change it. | On device by default: free, and your history stays on your computer. With Embed in the cloud on: Panda Cloud or your own Together.ai key. |
| Image generation | Creates and edits images in a chat. | Panda Cloud, or your own Together.ai or OpenAI key. Codex chats use your ChatGPT plan instead. |
Image generation is off until you turn it on from the Tools menu in the chat bar. Voice has its own settings page, see Voice.