---
title: "Playground"
subtitle: "Run a prompt on several models or settings side by side."
---

Use the playground to try a prompt on different models, settings, or versions before you commit it. A playground holds up to 8 prompt cards, runs them on the same inputs, and saves automatically.

## Open a playground

- **Gateway** › **Playground** opens your last playground, or creates one. Each playground is a tab. Click **New playground** (the icon after the tabs) to start another.
- **From a prompt:** in the prompt editor's side panel, click **Open in playground**, or press **P**. A new playground opens with that version's messages, model, and variables.
- **From a span:** in **Logs**, open a span and click **Run**, or press **P**. A new playground opens with the span's messages, model, and settings. See [Logs](/docs/documentation/features/observe/explore-spans).

## Compare models side by side

<Steps>
<Step title="Add a prompt card">

Click **+** to the right of the cards. The new card copies the active card's model and settings.

</Step>
<Step title="Load the prompt">

In the new card's footer, open the prompt dropdown (it shows **New prompt**) and pick your prompt. The card loads its messages, model, and variables. To write messages by hand instead, use **+ Message**.

</Step>
<Step title="Pick a model for each card">

Choose a model from the dropdown at the top of each card. Leave everything else the same, so the model is the only difference.

</Step>
<Step title="Run">

Click **Run all** (**⌘ ↵**), or a card's own **Run**. Each output lists its input and output tokens, cost, and latency.

<img src={"/docs/images/prompts/playground-compare-light.png"} className="block dark:hidden" alt="A playground with two prompt cards running the same Support reply prompt, one on gpt-5.5 and one on claude-sonnet-4-5, with each reply and its tokens, cost, and latency below" />
<img src={"/docs/images/prompts/playground-compare-dark.png"} className="hidden dark:block" alt="A playground with two prompt cards running the same Support reply prompt, one on gpt-5.5 and one on claude-sonnet-4-5, with each reply and its tokens, cost, and latency below" />

</Step>
</Steps>

## Card controls

| Control | What it does |
| --- | --- |
| Model dropdown | Sets the model this card runs. |
| **Configure model settings** (sliders icon) | Temperature, max tokens, and the model's other parameters. |
| **Configure variables** (braces icon) | Values for the card's `{{variables}}`. |
| **Tools** (wrench icon) | Functions the model can call. See [Tool calling](/docs/documentation/features/prompt-management/tool-calling). |
| **More prompt actions** (**⋯**) | **View code**, **MCP**, **Select preset**, **Override request parameters**, **Copy config to all**, and **Clear messages**. |
| **Remove prompt card** (trash icon) | Removes the card. |
| **Expand prompt card** | Opens the card full width. |

To add, edit, or delete the presets that **Select preset** applies, click **Presets and schemas** at the top right and choose **Presets**. A preset saves model settings and variable values. Choose **Schemas** to manage saved schemas for [structured output](/docs/documentation/features/prompt-management/structured-output).

## Save your changes to a prompt

- **Card linked to a prompt** (its footer shows the prompt's name): edits save to that prompt's draft as you work. Click **Commit** to save them as a new version, then deploy it from the prompt editor. See the [Quickstart](/docs/documentation/features/prompt-management/prompt-management-quickstart).
- **Card not linked** (its footer shows **New prompt**): click the pencil to name it, then click **Create** to save it as a new prompt.

## Legacy playground

The previous playground is still available. Open the **New** menu at the top right and choose **Legacy**. Choose **New** to switch back.