Models & Modes
Pro ships with 12 LLMs in the model picker and 5 permission modes for the agent.
Models
Pick from the header dropdown. You can switch mid-chat without losing history.
| Model | Strengths | Relative cost |
|---|---|---|
| Claude Opus | Highest quality, multi-step reasoning, long refactors | Highest |
| Claude Sonnet | Balanced; most turns go here | Moderate |
| Claude Haiku | Fastest, cheapest. Great for single-tool edits | Lowest |
| GPT-5 | OpenAI’s flagship; strong at structured code | High |
| GPT-5 Mini | Cheaper GPT alternative | Low |
| Gemini 2.5 Pro | Strong on vision and long context | Moderate |
| DeepSeek V3 | Cost-effective reasoning model | Low |
| Llama 4 | Open-weights alternative | Low |
| Mistral | Fast European option | Low |
| Kimi | Long-context alternative | Low |
Cost is deducted from your account after each turn; the header shows the running total for the current chat. Hover the number for a per-turn breakdown.
Choosing a model
- Big refactor touching many blueprints or files: Opus. The planning is worth the cost.
- Edit one blueprint, add an input action, tweak a material: Sonnet or Haiku.
- Vision-heavy prompt (you’re pasting screenshots): Opus, Sonnet, GPT-5, or Gemini. The cheaper open models don’t always handle images well.
- Trying to save credits: Sonnet or Haiku on simple jobs; Opus only when the task needs deep reasoning.
Modes
Modes control how aggressive the agent is with tool calls. Switch from the footer.
| Mode | Behavior |
|---|---|
| Default | Ask before destructive tools. Safe default. |
| Accept Edits | Auto-accept file edit operations. Skip the permission card for every .uasset write, but still ask for delete / rename / move. |
| Plan Mode | Only plan, never execute tools. Agent describes the change but never touches anything. |
| Don’t Ask | Never prompt; just deny anything not pre-approved for the session. |
| Bypass | Run everything. Power-user mode; review your diffs. |
When to use each
- Default is where you should start. Safe, forgiving, good for new users.
- Accept Edits once you trust a chat; switch to this so you don’t click “Allow” 50 times on a big refactor.
- Plan Mode when reviewing a proposed change before letting the agent near your project.
- Don’t Ask for running hands-off while you do something else. If the agent wants something it hasn’t been approved for, it fails instead of prompting.
- Bypass only when you really know what you want. The 209 tools include destructive ones (
delete_asset,write_file); Bypass lets them fire without asking.
Cost controls
- Switch models. Haiku is an order of magnitude cheaper than Opus for the same number of turns.
- Start fresh chats for unrelated tasks. Long history gets re-sent each turn, which inflates cost.
- Use Plan Mode when you’re just exploring ideas. No tool calls means no tool cost.
- Disable extensions you don’t need. Each extension sends additional context with every message.
Last updated on