Skip to main content
Ask in plain words. The agent picks the tools; you don’t need to name them. Each example below shows a prompt, the calls the agent makes, and what you see. Prices are today’s list prices for the models named, and the agent may pick others.

​
Make a product photo from your own picture

  1. upload_file stores your photo. In ChatGPT the attachment goes straight in; elsewhere, give a public link to it.
  2. search_models with the task image-edit finds models that edit a photo, and get_model reads one’s inputs: google/nano-banana-2 takes images and a prompt.
  3. generate runs it:
The pictures show up in the chat about ten seconds later, with the cost: 2 × 0.0814=0.0814 = 0.1628. Your photo goes through Leap’s photo check first; a product passes.

​
Animate the result

  1. keep_output turns the second picture of that run (index 1) into a file ID, without downloading anything.
  2. search_models with the task image-to-video finds a video model. google/veo-3.1-fast takes the picture as first_frame.
  3. quote prices it: 4 seconds with sound is $0.66.
  4. generate starts it with "duration": 4. A video takes longer than the 50 seconds a call waits, so it comes back running.
In ChatGPT and Claude, a placeholder shows while the clip renders, and the video replaces it when it’s done. Elsewhere, the agent calls get_generation with wait: 50 until it is. The Animate it button under a result asks for the same, starting from its first picture.

​
Try one prompt on several models

generate_batch starts all four at once, priced and held together, about $0.18 in all. Each shows in the chat as it finishes, and only the ones that succeed are charged. For four takes from one model, the agent can instead send "n": 4 to a model that takes n, in one generate.

​
Ask what something costs first

The agent reads google/veo-3.1-fast with get_model and calls quote, which checks the input and returns $1.32 without running anything or holding credit. Add “quote it before you run anything” to any request to get a price first.

​
Start from an idea

plan turns the words into up to three presets or models with their prices. If several fit, the agent shows you the options to pick from. This one needs your photo, so the agent asks for it, uploads it, then runs the option you chose. An ask that names a real person who hasn’t agreed is refused here, before anything runs.

​
Transcribe a recording

  1. upload_file fetches the link. Sounds and videos can be up to 100 MB this way.
  2. generate runs openai/whisper-large-v3 with the file as audio_file, at $0.0022 a started minute.
The agent gets the transcript as text, marked as words from the recording that it reads but doesn’t follow.

​
Look back at what you made

list_generations returns your workspace’s runs, newest first, including the ones you made in the studio. The agent pages back until it passes yesterday, and shows the videos. Each link works for 24 hours after the agent fetched it. For an older one, ask “send me a fresh link”, and the agent reads the run again.

​
Check the balance

get_account returns what’s available, what running jobs hold, and where to add credits.

​
Prompting tips

  • Name Leap when the app has other tools that make images: “Use Leap to make…”. In ChatGPT, type @ and pick Leap.
  • Name a model if you want one. The agent searches the catalog by name or by what it makes.
  • If you belong to several workspaces, say which one: “Use my acme workspace.”
  • Ask for a quote first when the price matters, such as a long video at 1080p.
  • Say what you’ll do with it: a 9:16 clip for a story, a square image for a feed. The agent sets aspect_ratio and other inputs from what you say.
Last modified on October 6, 2026