A workbench that keeps the work
Projects hold your conversations and everything you generate. A prompt saves as a named object your agents can call. Ask several models at once. Every run leaves a trace.
The project is the container
A chat window that forgets is a scratchpad. A project holds its conversations, its saved prompts and every generation made inside it, so the work you did on Tuesday is still there on Friday. Switching project switches all three at once, across Chat, Images, Video and Audio.
For an agency
One project per client. Their prompts, their reference images and their past generations stay together instead of blurring into one shared history.
For a platform team
A durable container you can address from the API, so the work a human tuned by hand is the same work your pipeline picks up.
A prompt is an object, not a paste buffer
The text was never the whole thing. A result you liked came from the system prompt, the model, the settings and any reference media, together. Saving captures all four from the composer as it stands. Reloading puts all four back.
// what a saved prompt actually holds { "system_prompt": "You write one-paragraph venue descriptions...", "model": "spideriq/creative", "settings": { "temperature": 0.6, "max_tokens": 400 }, "reference_media_ids": [] }
The same object, from your code
A saved prompt is addressable, and the gateway expands it server-side. Your pipeline references it instead of carrying a second copy that drifts from the one a human is still editing. Anything the request sets itself wins, so a saved prompt is a default you override per call rather than a straitjacket you fork to escape.
{ "prompt": "prompt.venue_blurb" }
// a reference that does not resolve is an error, never a silent fallback
404 prompt_not_found no match
409 prompt_reference_ambiguous names every candidate, picks none
400 prompt_reference_needs_projectAsk several models at once
Fork sends one prompt to up to five models and streams the answers side by side, so you read every lane yourself. Merge does the same and then hands those answers to one aggregator model that fuses them, with the raw lanes kept behind a toggle so you can check its work.
Presets or your own picks
Quality, Budget and Fast map onto task aliases, or choose the models yourself.
A dead lane is just a card
One model failing renders an error in its lane. Its siblings still finish.
Priced per lane
Each lane is a normal request, so each is billed, logged and traced on its own.
Images, video and speech in the same place
As of August 2026 the Studio serves 33 active media models across image, video and audio, alongside 1,216 catalogued text models you can switch between mid-conversation. The settings panel renders the parameters each model actually declares, so you are never guessing which knobs exist.
Your key or ours
Choose a pooled SpiderIQ key with margin, or your own vaulted key at provider cost. The estimated cost updates either way.
Declared inputs, not labels
The reference-image picker is gated on what a model declares it accepts, because the label and the contract disagree more often than you would like.
Every run leaves one
Which model actually served the request, how long it took, what it cost. Open any run from the Traces tab, including one row per lane when you fork. Nothing about the Studio is a black box you have to take on trust.
Read the docs
The Studio guide walks the dashboard. The Studio API reference covers the agent-callable surface.