# Models

> Pick the AI model behind your agent and balance answer quality, speed, and credit cost.

The model is the AI that writes your agent's replies. Some models give deeper, more careful answers. Others are faster and cheaper. Choosing a model is about finding the right balance for your customers and your budget.

**Where to find it:** open your agent, go to **Playground**, and look at the **Model** card on the **Overview** tab.

## How it works

Your agent uses one model for all its channels. Like everything else in the Playground, the model is part of your **draft**. Customers keep getting the old model until you publish.

Each reply your agent sends costs credits. Credits are what messages cost in Hourzero, and your plan includes a monthly amount. The model you choose sets the cost of each reply, from 1 to 7 credits.

1. **Choose**: Pick a model in the Playground.
2. **Save**: Select Save draft.
3. **Test**: Ask your usual questions and compare.
4. **Publish**: Publish each channel to use the new model.

New agents start on **Auto**, which costs 1 credit per reply. It's a good choice for most businesses.

## Change the model

1. Go to **Playground** and stay on the **Overview** tab.
2. Under **Model**, select the current model name. A list opens, grouped by provider: **OpenAI**, **Anthropic**, **Google**, **xAI**, **Moonshot AI**, **Z.ai**, **DeepSeek**, and **Mistral**.
3. Optional: type in **Search models...** to find one by name.
4. Hover over a model to see a short description and its price, for example **Pricing: 3 credits per response**.
5. Select the model you want.
6. Select **Save draft** at the bottom of the settings panel.
7. Test the agent in the preview with the questions your customers ask most.
8. When you're happy, select **Publish** for each channel that should use the new model.

> [!IMPORTANT]
> Changing the model changes your draft only. Publish each channel, one at a time, to use the new model there. Until then, live customers get the model that was published before.

## How to choose

You don't need to understand how AI models work to choose well. Think about three things:

- **Quality**: how well the agent handles tricky, multi-step, or unusual questions.
- **Speed**: how fast replies arrive. Lighter models usually answer faster.
- **Cost**: how many credits each reply uses.

Some practical starting points:

- **Answering common questions from your help center**, like shipping times or return policies: stay on **Auto** or another 1-credit model. These handle everyday questions well.
- **Walking customers through multi-step tasks**, like a return that uses actions and procedures: try a 2- or 3-credit model if the agent misses steps or gets confused.
- **Complex, high-stakes conversations**, like detailed product advice or sensitive account issues: test a 4-credit model and check whether the better answers are worth the extra cost.

> [!TIP]
> Test a new model with the same set of questions you used before. Keep a short list of 5 to 10 real customer questions, including a few hard ones. If the answers aren't clearly better, stay with the cheaper model.

## Models and credit costs

These are the models you can choose from today, grouped by how many credits each reply costs.

| Credits per reply | Models                                                                                                                                         |
| ----------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
| 1                 | Auto, GPT-6 Luna, Gemini 3.5 Flash Lite, GLM 5.3 FlashX, DeepSeek V4 Flash Vision Exp, DeepSeek V4.1 Flash 1017, DeepSeek V4.1 Flash 1017 Fast |
| 2                 | GPT-6.1 Terra, Claude Haiku 5.5, Gemini 3.8 Flash, GLM 5.3, DeepSeek V4 Pro, Mistral Large 4                                                   |
| 3                 | GPT-6.1 Sol, Claude Sonnet 5.5, Gemini 3.1 Pro Preview, Gemini 4 Argon, Grok 4.7, Kimi K3                                                      |
| 4                 | Claude Opus 5.5, Kimi K3 Fast                                                                                                                  |
| 7                 | Claude Fable 5.1                                                                                                                               |

The model picker always shows the current price when you hover over a model. If it ever differs from this table, the picker is right.

## What a model choice costs over time

Small differences per reply add up. Here's what 1,000 agent replies cost at each price:

```text
1-credit model:  1,000 replies × 1 credit  = 1,000 credits
3-credit model:  1,000 replies × 3 credits = 3,000 credits
7-credit model:  1,000 replies × 7 credits = 7,000 credits
```

A few other things also use credits:

- **Playground tests.** Your test messages cost the same as real customer messages.
- **Skills that create files.** Some [skills](https://hourzero.dev/docs/ai-agent-management/skills), like making a PDF, run a small program in the background. That time costs 1 extra credit for every 2 minutes it runs, or part of 2 minutes.

> [!WARNING]
> Switching to a more expensive model raises the cost of every reply on every channel you publish it to. Before you switch, check your monthly credits on the [Usage](https://hourzero.dev/docs/workspace-management/usage) page. To cap spending for one agent, set a [weekly credit spend limit](https://hourzero.dev/docs/ai-agent-management/guardrails).

## Common questions

### Can I use a different model on WhatsApp than on my website?

No. The model is shared by all channels of an agent. You can still change how the agent behaves on each channel with channel-only instructions in the Playground.

### Does the model change what my agent knows?

No. Your agent's knowledge comes from your [data sources](https://hourzero.dev/docs/ai-agent-management/data-sources), whichever model you choose. A stronger model may use that knowledge more carefully, but it doesn't add new facts about your business.

### How can I compare two models before publishing?

Test both in the Playground with the same questions. For a more thorough check, run a [benchmark](https://hourzero.dev/docs/ai-agent-management/benchmarks) to compare your draft with what's live.

## Next steps

- [Playground](https://hourzero.dev/docs/ai-agent-management/playground): Test your agent and publish your changes.
- [Usage and credits](https://hourzero.dev/docs/workspace-management/usage): See how many credits your workspace has used.
- [Benchmark Agent](https://hourzero.dev/docs/ai-agent-management/benchmarks): Compare answers from your draft and your live agent.
- [Guardrails and abuse protection](https://hourzero.dev/docs/ai-agent-management/guardrails): Set a weekly credit limit for this agent.
