Skip to main content
The playground runs any model in the catalog from your browser. After each run it shows what the call cost, how long it took, the time to the first token (for chat models; speech models show the time to first audio), the tokens used (with the cached and reasoning share when the model reports them) and the request id to quote to support.

Open it

  • Dashboard → Models, then Try on any model, or Open the playground at the top of the list.
  • Open in playground on a model’s own page.
  • Straight to a model: /dashboard/models/playground?model=<model id>.
Pick another model from the search box at the top of the page. Each run is billed to the workspace and environment selected in the header, through the environment’s default API key.

What each model kind gets

The controls come from what the model takes, as the catalog lists it.

Chat and vision

  • Replies stream in as they are written, in a conversation you can keep going.
  • A system prompt, max tokens, temperature and the model’s other sampling settings.
  • Tools: paste function definitions as JSON. The model’s calls show in the reply; type each result back and send it to continue.
  • Structured output: ask for JSON, or give a JSON Schema. The reply is checked against the schema when it lands.
  • Reasoning shows folded above the reply, and you can set the effort where the model supports it.
  • Images: attach pictures to a vision model; they are sent inline as base64.

Image

A prompt, the model’s options (size, aspect ratio, seed and so on) and an input image where the model edits one. A model priced by its options shows each option with its price.

Video

A prompt and a length. A video runs as a job: the page shows whether it is queued or generating and for how long, then plays it with a download. You can leave the page; the run stays in Recent runs.

Speech

Text and a voice. Playback starts with the first audio, then the clip is yours to replay or download as WAV.

Transcription

Choose an audio file (MP3, WAV, FLAC, M4A or WebM, up to 25 MB) or record one, then transcribe it. The transcript shows the length of audio billed and the language heard.

Embeddings

One input per line and, where the model supports it, a shorter number of dimensions. Each vector shows its length and first numbers, with how alike the inputs are to each other.

Compare

Compare at the top of the page sends one prompt to up to four models of the same kind at once, each answer with its own cost and latency, and for chat models its time to first token.

Get code

Get code turns your last run (or the settings as they stand) into the same setups the model’s page shows, in this order:
  • One line for your agent. Pick what you want: Add it to my project (your coding agent adds the SDK, your key and one function, then calls it once), Run my coding agent on it (text models that call tools: the agent points itself at the model), or Use it in an app (Open WebUI, Dify, LlamaIndex or LiveKit, by the model’s kind, in two or three steps). Signed in, the line carries a one-time link that makes a key for your agent.
  • Set it up by hand. Each agent’s own settings: Claude Code, Codex, OpenCode, Cursor, and Cline, Roo Code or Kilo Code, each set to the model and its context window.
  • Call it from code. The OpenAI SDK in Python and TypeScript, the Anthropic SDK in Python and TypeScript (base URL https://api.runcrate.ai), the Vercel AI SDK and LangChain, then cURL last. Each streams the reply and carries the run’s own settings: its tools, image, reasoning effort or JSON schema. Video shows the whole job: start it, check it, download it.
Each reads your key from RUNCRATE_API_KEY. Create one in API keys.

Recent runs

Every run you make is kept for 30 days under the page, with what it cost and how long it took. Open one to see it again with its request filled in, and run it again. Pictures, audio and attachments are not kept with a run; finished videos can be fetched for a while after they are made.