Start with the job, not the tool
Five things people actually do here, in the order they tend to do them: make one piece, work out which model to trust, make the next twenty the same way, drive it from code, and put everyone on one balance. Pick the one that sounds like your week.
Script it, storyboard it, animate it, voice it — in one thread
Write the script in chat with whichever model you like. Ask for the scene and it arrives as an image that matches what the two of you just agreed. Animate the frame that works. Add the voiceover. Nothing gets exported, re-uploaded, or re-explained on the way.
- The character you described in message two is the same character in the video
- The thread survives a change of model, not just a change of tab
- Every asset lands in the workspace, not in your downloads folder
Send one prompt to five models. Keep the one that won.
Finding out what a model can do should not cost a subscription. Send the same prompt to as many as five at once, put the results beside each other, and have a model review them if you would rather not judge your own work. Then spend the rest of the month on the winner.
- Up to five models per comparison, one prompt, one balance
- An optional reviewer model scores the results for you
- Costs credits, not another card on file
The chain you just did by hand becomes one run
The steps you took for the first piece are a graph. Save it, and the next twenty pieces are that same graph with different inputs — including the parts that were never a model call: the crop, the transcode, the script that renames things, the search you did first.
- Seven kinds of step and 31 named transforms — code, HTTP, web search, document parsing, video and audio work, fan-out over a list
- Per-step artifacts, so when step five fails you can look at what step four made
- Hand the run to someone who does not write code
You don't write the orchestration
One POST carries the whole graph: model calls, media generation, transforms, code, HTTP, web search, fan-out over a list. The alternative is your own queue, your own retries, your own intermediate storage and transcode — weeks of it, and then maintenance forever.
- The canvas and the API are the same thing: read a graph as JSON, or paste a curl command to fill in an HTTP step
- When a multimodal run fails at step five, the intermediates are images and video — you cannot read those in a log
- Swap the base URL and keep your SDK; the same key now also reaches image, video, audio, music and 3D
POST /v1/pipelines
{
"name": "Scene → clip → cut with the voiceover",
"definition": {
"inputs": [
{ "name": "brief", "type": "text", "label": "Scene brief" },
{ "name": "voice", "type": "audio", "label": "Voiceover" }
],
"steps": [
{ "id": "scene", "type": "media", "model": "nano-banana-2",
"input": { "prompt": "${input.brief}" } },
{ "id": "clip", "type": "media", "model": "veo-3.1-fast",
"input": { "prompt": "${input.brief}",
"image": "${steps.scene.output.url}" } },
{ "id": "final", "type": "video.add_audio",
"input": { "video_url": "${steps.clip.output.url}",
"audio_url": "${input.voice}" },
"params": { "mode": "replace" } }
],
"output": { "video": "${steps.final.output.url}" }
}
}One balance, one workspace, five roles
A household, a two-person studio or a ten-person agency run the same mechanic: one prepaid balance instead of seats bought ad hoc, and roles so the person who pays does not have to be the person who reads the work.
- Five roles — owner, admin, developer, billing, viewer. A billing-only member never sees Studio data
- Shared chats, shared runs, shared templates
- Caps per API key: requests per minute, allowed models, and a credit budget per hour, day, week or month
Pick a job and try it
Free trial credits, no card. Enough to run a chain end to end and see for yourself whether the thread holds when the modality changes.