> ## Documentation Index
> Fetch the complete documentation index at: https://docs.genieforge.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Choosing a chat model

> When to use each chat model in Build and Chat, including cost tiers, thinking, and HIPAA-safe options.

# Choosing a chat model

Pick a model in the chat composer, or set **Default Chat Model** under **App Settings → Overview**. The picker shows a relative cost cue next to each name (`$` through `$$$$`). Those signs compare models to each other; they are not dollar quotes. Actual credit use still depends on how much work a turn does. See [Controlling credits](/account/controlling-credits).

You can also pick **Thinking** effort separately for models that support it. Higher thinking usually means slower, more careful turns and more credits.

## Quick pick

| Goal                                          | Start with            | Cost   |
| --------------------------------------------- | --------------------- | ------ |
| Everyday Building and Chat (default path)     | **Claude Sonnet 4.6** | `$$`   |
| Hard architecture, messy bugs, deep plans     | **Claude Opus 4.6**   | `$$$$` |
| Fast, cheap answers or light classification   | **Claude Haiku 4.5**  | `$`    |
| Cheap HIPAA-friendly alternative to Haiku     | **Kimi K2.5**         | `$`    |
| OpenAI flagship for Build or a Second opinion | **GPT-5.6 Sol**       | `$$$`  |
| Balanced OpenAI                               | **GPT-5.6 Terra**     | `$$`   |
| Cheapest OpenAI                               | **GPT-5.6 Luna**      | `$`    |
| Non-HIPAA alternative with strong reasoning   | **Grok 4.5**          | `$$`   |

If you are unsure, stay on **Claude Sonnet 4.6**. GenieForge may use a cheaper model for small, scoped Build edits; you do not need to micro-manage every tweak.

## Claude family

### Claude Sonnet 4.6 (`$$`)

Best default for Building pages, tools, migrations, and everyday Chat. Strong at following GenieForge building steps and recovering from ordinary errors. Use Thinking (Low or Medium) when a plan needs more care; reserve High for stubborn problems.

### Claude Opus 4.6 (`$$$$`)

Use when Sonnet keeps missing the mark: ambiguous requirements, large refactors, subtle permission or data bugs, or plans that span many surfaces. Expect higher credit burn. Switch back to Sonnet once the hard part is unstuck.

### Claude Haiku 4.5 (`$`)

Use for light Chat, quick Q\&A about the app, or high-volume work where speed and cost matter more than deep coding. Prefer Sonnet for serious Building. Haiku is a good fit for some sub-agent or background-style tasks when you pin a cheaper model on purpose.

## Other models

### Kimi K2.5 (`$`)

A lower-cost option that stays available when your app needs HIPAA-eligible models. Useful for debugging and agentic edits when you want something cheaper than Sonnet and stronger than a pure Q\&A pass. Still prefer Sonnet for large greenfield builds.

### GPT-5.6 Sol, Terra, and Luna (`$$$` / `$$` / `$`)

OpenAI’s flagship (Sol), balanced (Terra), and cost-efficient (Luna) tiers. Good when you want a different model family for Building, Chat, or a [Second opinion](/building/second-opinion). Sol is the capable end of that family; Luna is the cheap end. These are not offered on HIPAA apps unless support has enabled an OpenAI path for your account.

### Grok 4.5 (`$$`)

An xAI option with reasoning on by default (you can still choose Low / Medium / High). Useful as a non-HIPAA alternative for analysis or a Second opinion. Not available for HIPAA apps.

## HIPAA apps

When HIPAA mode is active for an app, the picker only lists models that are allowed for that arrangement (Claude Sonnet, Opus, Haiku, and Kimi today). GPT-5.6 and Grok stay hidden unless support has enabled them for your account. For HIPAA setup itself, see [HIPAA and regulated health data](/account/hipaa-and-regulated-data) and contact **[support@genieforge.ai](mailto:support@genieforge.ai)**.

## Thinking effort

For models that support it, **Thinking** is a separate control next to the model:

* **Off** (when available): fastest and usually cheapest for straightforward edits
* **Low / Medium**: better plans and fewer blind retries on medium work
* **High**: reserve for hard debugging or careful design

Grok always runs with reasoning; there is no Off.

## Where to set it

* **This chat**: model control in the composer
* **App default**: **App Settings → Overview → Default Chat Model** (conversations that do not override inherit this)
* **Agents / sub-agents**: each profile can pin its own model when you want Chat or automated work on a cheaper or stronger lane

## Related

* [Using the AI chat](/building/using-the-ai-chat)
* [Second opinion](/building/second-opinion)
* [Controlling credits](/account/controlling-credits)
* [Credit usage in an app](/account/credit-usage)