# Models

> Which models power Glow's AI steps, how to choose between them, and why you do not need an API key to get started.

Glow's AI steps run on Gemini models, on infrastructure we host and pay for. You do not supply an API key, manage rate limits, or hold a contract with a model provider.

That is handled for you, and usage is billed through your Glow plan.

## Which step chooses a model

Only one of the two AI steps lets you pick a model.

| Step                                      | Model selection                                              |
| ----------------------------------------- | ------------------------------------------------------------ |
| [AI Prompt](/build/ai-features/ai-prompt) | None. Runs on a managed default tuned for single-call tasks. |
| [AI Agent](/build/ai-features/ai-agent)   | **Brain (model)** dropdown in the App drawer.                |

[AI Data Transform](/build/ai-features/ai-transform) runs on the same managed default as AI Prompt.

If you want control over the model, use the AI Agent step.

> The field is labelled **Brain (model)** on the canvas. New models appear in
> the dropdown as they ship, so it always shows the current list.

## Choosing a model

Model names change by generation, but they sort into three families and the trade-off between them does not move:

**Flash** is faster and cheaper. Use it for classification, routing, extraction, short summaries, and anything running at high volume.

**Pro** gives stronger reasoning over long or ambiguous input. Use it for multi-step agent work, complex tool use, careful document analysis, and drafting.

**Flash Lite** variants trade further capability for speed and cost. They suit narrow, well-specified tasks such as single-label classification.

A new AI Agent step starts on a **Pro** model, which is the safe default rather than the cheap one. For classification, routing and extraction, switch it to Flash and check the output against a handful of real cases. On a high-volume workflow that is usually the single biggest saving available. Move back to Pro when you can point at a specific output the smaller model got wrong.

### By Task

| Task                                    | Start with |
| --------------------------------------- | ---------- |
| Classification, routing, tagging        | Flash Lite |
| Summarization, field extraction         | Flash      |
| Multi-step agent reasoning and tool use | Pro        |
| Long or unstructured document analysis  | Pro        |

### By Volume

High-volume workflows amplify both cost and latency. A workflow running 1,000 times a day makes the Flash/Pro choice roughly a thousand times over. Measure on a representative sample before committing a Pro model to a hot path.

## Setting it up

1. Select the **AI Agent** step on the canvas to open the App drawer.
2. Open the **Brain (model)** dropdown.
3. Pick a model.

Different steps in the same workflow can use different models. A common pattern is a cheap model for an initial triage decision and a stronger one only on the branch that needs it.

## API keys

You do not need one. Every Glow AI step runs on managed infrastructure, and there is no key to configure before your first workflow.

If you would rather use your own provider, the Apps menu carries connectors for [OpenAI, Anthropic, Mistral and the rest](/build/ai-features/overview#the-providers-you-can-connect). Those take your key and bill to your account with that provider.

## What's Next?

- Write single-call prompts with the [AI Prompt](/build/ai-features/ai-prompt) step.
- Build autonomous, tool-using logic with the [AI Agent](/build/ai-features/ai-agent) step.