Choose Bali's AI model, reasoning and behavior
You can tune how Bali works on each project with three settings:
- Model: the AI model behind Bali’s conversations
- Reasoning: how much thinking to request from models that support it
- Behavior: how independently Bali works and how often it checks in with you
All three are saved per project, so you can use a fast model for a quick prototype and a stronger one for a complex game.
Where to change them
Section titled “Where to change them”- In the chat: click the model chip next to the + button under Bali’s chat box. The menu has Model, Behavior and Reasoning options.
- In Project Settings: click the gear icon in the top-right toolbar and open AI Settings. See Settings.
Models
Section titled “Models”The model menu groups models by category and shows each one’s provider and a short description. The list changes as new models are released, and some models may not be available on every account. If a model shows a label next to its name, it isn’t available to you right now.
Recent additions include:
| Released | Models |
|---|---|
| October 2026 | GPT 6.1 Sol |
| September 2026 | GPT 6 Astra, GLM 5.3, Gemini 3.7 Flash, Grok 4.6 reasoning |
| August 2026 | GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, GLM 5.2, Kimi K3 |
| June 2026 | Gemini 3.5 Flash |
Studio always shows the current list, so check the model menu for what’s available to you. What’s new announces new models as they arrive.
Choosing a model
Section titled “Choosing a model”- Fast models (often labeled Flash, Lite or similar) are great for quick questions, small edits and brainstorming. They usually use fewer Sparks.
- Stronger models handle complex coding, large refactors and tricky bugs better, but usually use more Sparks per request.
- Efficiency varies a lot between models. For example, GPT 6.1 Sol performs close to GPT 6 Astra on coding and complex work while using around 80% fewer Sparks for comparable token usage.
- Not every model can see images. If you attach an image and the selected model can’t read it, Studio shows a warning on the image: “This image won’t be included in the context because [model] doesn’t support image inputs.” Switch to a model that supports images if Bali needs to see it.
If a model isn’t working well for a task, switching models mid-project is fine. Bali keeps your project and chat history.
Watch: Trade-offs between speed and quality (0:45, from the masterclass Designing Game Systems That Feel Good)
Reasoning
Section titled “Reasoning”Reasoning models can spend more time thinking through a problem before they answer. Default (Medium) suits most work. The available levels depend on the model, for example None, Low, Medium, High and Extra High.
- Use higher reasoning for complex systems, hard bugs and planning large changes.
- Use lower reasoning for quick edits and simple questions. It’s faster and uses fewer Sparks.
If the selected model doesn’t support reasoning, the menu shows Reasoning: Not supported.
Behavior
Section titled “Behavior”Behavior sets how much Bali does on its own before checking in with you. For small tasks, every behavior focuses on exactly what you asked. The differences show up on medium and large tasks.
| Behavior | In short | What Bali does |
|---|---|---|
| Autonomous | Rapid, independent action | Makes changes quickly, with minimal interruptions or confirmations. Asks only about critical decisions. |
| Collaborative | Balanced collaboration and autonomy | Checks in periodically, providing updates and confirming significant actions. |
| Cautious | Careful, user-guided actions | Frequently asks for confirmation and clearly explains each step before making changes. |
| Creative | Exploratory and innovative | Proactively suggests new ideas and experimental changes beyond your explicit instructions. |
When to use each behavior
Section titled “When to use each behavior”Pick the behavior that matches how clear your goal is.
Autonomous
Section titled “Autonomous”Use Autonomous mode when you already know what you want and can describe it clearly. It’s useful when:
- You have a specific goal
- You can describe the expected result
- You want Bali to make progress without many follow-up questions
- You’re comfortable reviewing and debugging the result
Autonomous mode works best for advanced or highly descriptive users.
Collaborative
Section titled “Collaborative”Use Collaborative mode when you’re still exploring ideas or want Bali to confirm the direction before making major changes. It’s useful when:
- You’re unsure what direction to take
- You want options before committing
- You want to reduce debugging prompts
- You want Bali to ask clarifying questions
- You’re designing a system for the first time
Collaborative mode is great when you want Bali to help shape the idea with you.
Cautious
Section titled “Cautious”Use Cautious mode when you want to approve each step. Bali explains what it plans to do and asks for confirmation before making changes. It’s useful when:
- You’re working on a fragile or complex part of the game
- You want to learn how the game works as Bali changes it
- You’re close to publishing and want to avoid surprises
Creative
Section titled “Creative”Use Creative mode when you want ideas, not just execution. On bigger tasks, Bali proactively suggests new ideas and experimental changes beyond what you asked for. It’s useful when:
- You’re brainstorming mechanics, story beats or visual directions
- You want variations to choose from
- The game feels flat and you want fresh angles
Switch as your project matures
Section titled “Switch as your project matures”Many creators use Creative or Autonomous while exploring an idea, then switch to Cautious once the game works and they want to protect it. Watch at 22:06 or 56:49
Watch: Pick a behavior for each phase of your project (1:10, from the masterclass Designing Game Systems That Feel Good)