Playground

Try any catalog model without writing code. The portal's Playground streams real responses through the same live endpoint your app will use.

Controls

Model picker, a dropdown of every model on the Models & pricing page, with maker logos, and the context window and pricing for the selected model.
System prompt, a collapsible popover in the composer.
Parameters rail, temperature, max tokens, top-p, response Format (text or JSON object), and a plain Reasoning switch, on by default, for models that think (how reasoning control works).
Images, on vision models, attach up to 4 images of 2 MB each, or paste an image URL. If you switch mid-thread to a text-only model, the Playground removes the images from the history and shows a notice saying so.
Key picker, choose which of your API keys a run bills to, so Playground traffic appears under the key you expect in Usage.
Your real key and balance, runs bill your workspace at normal rates and show up in Usage like any other API call. With no key, or an empty balance, a banner names what's missing and links to the page to fix it.

Answers render as markdown. Nothing is stored: the conversation exists only in the browser tab, and closing the tab discards it.

Metrics on every response

The Last response card measures each response client-side:

Metric	Meaning
TTFT	Time to the first token of any kind, including the first reasoning token.
E2E	Full wall-clock time to the last token.
Tokens	`N in · M out` from the response's own `usage` block.
Speed	Output tokens over the generation window.
Cost	What that response cost, at rate-card prices.

A running session cost totals everything sent in the tab.

Reasoning

When a model streams its thinking, the response bubble shows a collapsed Thinking… section that fills in live. Once the answer starts, it becomes a Reasoning dropdown you can expand at any time. The trace is display-only: the Playground never sends it back as conversation context, so it does not add to the next turn's cost.

Errors

If an upstream model rejects a request (an unsupported parameter, reasoning that can't be disabled), the Playground shows the same error the API returns; it does not retry with a different configuration. That includes telling two 402s apart: an empty workspace balance and a key that has spent its budget are different problems, and the Playground names the one you have.