Playground

The Playground runs a model or a ready deployment with a realistic input. Runs are real requests: they are billed to the workspace and show in Usage under the Playground's own label. Owners, admins and developers run it; other roles read it.

• A conversation: each message sends the whole conversation, with the system message first. An optional system message says how the model should answer.
• Enter sends; Shift Enter adds a line. Each answer streams in as it is written, rendered as markdown, code colored.
• An image, where the model takes one: JPEG, PNG or GIF, sent inline with your next message, at most 2.5 MB. WebP is not read on providers' GPUs.
• Tools, as the OpenAI shape's tools array in JSON.
• Other model types take their own input: a file through /v1/uploads for audio, or the inputs a deployment declared.

A run carries the pool, pricing mode, region and fallback you choose, Auto first. A router can be chosen as the model, with its preference. The Playground never relaxes a requirement: when supply is Unavailable under it, the run says so instead of sending elsewhere.

The model that served it and the pool, the usage, and the speed where the gateway sent the times; a speed is never estimated. A market quote binds when the gateway accepts the request, and the run shows its rate and charge. A request a guardrail acted on says so.

When the workspace has no credit left, the run stops and a dialog offers to choose a plan or add credit; the draft is kept. When a budget's action pauses new requests, the dialog says to raise or clear the budget in Billing, or wait for the next month; a payment does not lift it.

Every run can be copied as code: the request with the model or deployment id and the requirement headers of the run filled in, where HIPS_API_KEY, HIPS_BASE_URL and HIPS_MODEL stand in for your values. No key is ever shown.

On a plan where request data may be used for training, the Playground says so beside the input, as every request of the workspace runs under the plan's policy.