Choose workspace models
Admins choose which AI model families each team can use, set fallbacks, and set how long chats are compacted.
The Models page decides which artificial intelligence (AI) models your workspace can use. Agents follow the same rules. Only admins can change the settings. Other members see them read-only.
Select Models in the settings sidebar. The page has two tabs, Access and Compaction.
Model families
Section titled “Model families”TulipFarm lists models by family. The exact version behind a family can change over time.
- Claude Haiku, Claude Sonnet, Claude Opus, Claude Fable
- GPT Luna, GPT Sol, GPT Astra
- DeepSeek Pro, DeepSeek Flash
- Grok, Grok Fast
- Kimi, Kimi Code
The models that people see in chat depend on your workspace settings. The default chat family is GPT Sol. A saved agent with no chosen family uses GPT Luna.
Access tab
Section titled “Access tab”On the Access tab, admins set these items for each model family:
- Which teams can use the family.
- A Fallback family. If a model is unavailable, chats switch to its fallback. You can also choose No fallback.
Open a family to see its details:
| Detail | What it is |
|---|---|
| Provider | The company that runs the model. |
| Current mapped version | The model version that the family uses today. |
| Input credits per million tokens | The credit cost of the text that the model reads. |
| Output credits per million tokens | The credit cost of the text that the model writes. |
Compaction tab
Section titled “Compaction tab”Compaction shrinks a long chat’s working context so that the chat can continue. The original transcript stays readable. These settings apply to every chat in the workspace.
| Setting | What it is |
|---|---|
| Compact automatically | Turns automatic compaction on or off. It is on by default and starts at 70% of the context budget. |
| Trigger | The share of the model context at which compaction starts. |
| Target | The share of the model context to shrink down to. |
| Summary model | The model that writes the summary: Claude Haiku, DeepSeek Flash, or Grok Fast. |
| Fallback summary model | Optional. TulipFarm uses it only when the summary model is unavailable. |
| Handoff instructions | Optional guidance for the summary, such as the audience or format. Up to 2,000 characters. |
When automatic compaction is off, people can still type /compact in a chat.
-
Change the settings on the Compaction tab.
-
Select Save compaction settings.