This update is about two things you notice in your first minutes: what you pay for, and how long you wait. A generation that fails now costs you nothing, a locked model tells you why, and replies start sooner and get to the point.
A generation that fails costs nothing
When an image, a picture edit, speech, a transcription or a translation fails on the side of the service that makes it, for example because it takes too long or returns an error, you are no longer charged for it. If you cancel one yourself, you pay only for what the service reports it had already done, and nothing if it reports nothing. A video that the service cancels is refunded in full. When nothing was charged, the error in the studios says so.
Know why a model is locked
A locked model in the model list now tells you why: your plan does not include it, or your plan's allowance and your wallet cannot cover even a short reply on it right now. Before, a model in the second case looked available until you sent a message. To go on, top up your wallet or pick another model.
Replies that start sooner
Behind every message, the AI also receives instructions about what it can do. Those instructions are now shorter, with nothing taken away, so the first reply in a new chat starts sooner. With Auto, among models of the same quality, ToRun now puts the ones that start answering quickly first, and it never trades quality for speed. On the Free plan, Auto answers chat messages with a model that starts its reply much sooner.
Replies that get to the point
Replies skip the introduction, do not repeat your question and do not end with a summary you did not ask for. When you ask for detail, a whole document, code, a plan or a translation, you still get all of it, and a greeting gets a friendly reply and a short question.
Also in this update
- Before changing a Canvas, the AI reads an outline of each file. When the servers were busy, that outline could come back empty. Now it is tried again, and if it is still incomplete, the AI searches the file instead, so its changes land in the right place.
- At a venue QR code that knows where it is placed, the venue assistant now answers "Where am I?" with the floor you are on and what is nearest, instead of turning it down.
- A short venue question that does not show its language is answered in the language of the conversation, with the directions in that language too.
- When the venue assistant's answer names a stand, asking to see it on the plan shows that stand, and "How do I get there?" gets directions to it. When no place has come up yet, the assistant asks where you want to go.