Agent Models
The Main Agent runs on a configurable model tier. Each task can use a different tier, and the agent supports per-step cost routing — heavier reasoning steps use a stronger tier while light bookkeeping steps fall back to a cheaper one.
Model Tiers
DeepSeek V4 Pro
Strong reasoning at low cost. A solid default for everyday creative tasks and tool calls.
Kimi K2.6
Long-horizon reasoning with a 262K context window. Best when the task spans many turns, large documents, or multimodal context.
Claude Sonnet 4.6
Maximum quality for complex planning, structured output, and customer-facing copy. Slower and more expensive — use when quality matters most.
The default for new tasks is Balanced II (Kimi K2.6). Workspace Owners and Admins can change the workspace default in Settings → Agent.
Per-Step Cost Routing
When cost routing is on, the agent picks the tier per step:
- Planning steps use the configured tier (typically Balanced II or Smart).
- Routine tool calls (extracting fields, formatting outputs, posting comments) drop to a cheaper option.
- High-stakes generation (final image / video creation, customer-facing copy) stays on the configured tier.
This keeps long tasks economical without sacrificing the quality of the steps that matter.
Choosing a Tier
| Task profile | Recommended tier |
|---|---|
| Generate a social media kit | Balanced I |
| Triage a Jira ticket and post a structured update | Balanced I |
| Plan a multi-week campaign with content calendar | Smart |
| Long-document analysis (briefs, research, transcripts) | Balanced II |
| Daily summary of a Monday board | Balanced I |
| Draft branded copy for a launch email | Smart |
| Multi-turn conversation with many attachments | Balanced II |
You can override the default tier per task in the task settings panel.
Switching Tier Mid-Task
If a task starts on Balanced I but the agent detects a step that exceeds the tier's reliability threshold, it can promote the step to a stronger tier automatically. This is logged as a tier switch event in the task timeline so you can audit the reasoning.
Related Pages
- Overview — what the Main Agent does
- Skills — tier recommendations vary by skill complexity
- Credits System — agent runs consume credits per token and per tool call