AI ASSISTANT

Say what you need. It works out the how.

Choosing between 46 models is a job in itself, and so is writing a prompt in the style each one responds to. Describe the outcome instead , “a product campaign with three moods and a short clip” , and the assistant lays it out for you.

It proposes. You decide.

The assistant never starts a run on its own. It shows you the setup it has in mind, which model does which step, and what the whole thing would cost. Nothing is charged until you press start. That is the point of it: the expensive part is not generating, it is generating the wrong thing five times.

It cannot invent a model

Whatever it proposes is checked on our side against the real catalogue before you see it. A model id that does not exist is refused , not quietly swapped for something similar. The reason is simple: if we substituted, something else would run than what was on your screen, at a different price, and you would have no way to know.

It knows how each model likes to be spoken to

These models genuinely differ. One reads a prompt as a shot brief and gets better the more you tell it; another wants one short line, subject first, and gets worse if you stack adjectives. A prompt that works beautifully on one can produce rubbish on the next. That knowledge was researched per model family, with the source noted , and the models we have not researched get no recommendation rather than a guessed one.

What it costs

The assistant runs on a paid language model, so a message costs a small number of credits, shown before you send it. If the call fails, those credits come back. Generating is priced exactly as it is everywhere else in the studio , the assistant gets no special rate, and it uses the same execution path you do.