Generating & credits

Every image, video and audio clip you make is a generation. Generations cost credits, and the price is always shown before you commit.

What a generation costs

The cost depends on the model, the resolution, the length and how many references are attached. The Generate button shows the exact figure for the settings in front of you — that number is what gets charged.

For some image models the length of the prompt counts too, because the provider charges for the words it reads. The figure on the button updates as you type, so a very long prompt shows its cost before you generate. Characters in Chinese, Japanese and Korean count for more than Latin letters, since each one carries more text for the model to read.

Counts assume the settings a new node starts with — GPT Image 2.5 Flare at 1K, medium quality, no references (2 credits), and Seedance 2.0 Fast at 480p for 4s (48 credits). Higher quality, larger sizes, longer clips and other models cost more. One credit is $0.01.

Comparing prices before you generate

The settings card in the inspector shows what the current settings cost next to a compare link. Open it and you get every model you can pick, priced for those same settings and sorted cheapest first.

Models do not all accept the same settings. Any model that cannot take the exact settings you are comparing at — no 480p, or a length it does not offer — is left out, and a line at the bottom says how many were hidden. Every price in the list is therefore for the same thing, so the cheapest row really is the cheapest. Widen the settings and the hidden models come back. Picking a row switches the model and keeps the price you were quoted.

The resolution and length controls at the top of the comparison change the settings themselves, not just the table, so you can leave it having chosen both a model and the settings you compared at.

Audio nodes compare the same way, within one kind — sound effects against sound effects, music against music. A speech model and a music model are not alternatives to each other, so they never share a table. Some audio models choose their own length; those have no length control to compare at.

Editing a clip instead of regenerating

Some video models edit an existing clip rather than making a new one. Pick one in the model row and the panel changes: you choose a source video — or connect one to the node on the canvas, which does the same thing — then either Edit it — describe the change in plain language, such as make it snow — or Extend it, adding 5 to 20 seconds to the end.

Setting a start frame works the same way on some models: when the output has to follow the frame you gave it, the ratio locks and the picker says so. Nothing is lost — the result matches the frame it continues from.

Generated clips are finished at a standard size for their resolution and ratio — a 480p 16:9 clip is always 854×480, whichever model made it. Some models return a few pixels more on one side; the small difference is rescaled away so that clips from different models, and a shot continued from another shot's last frame, line up on the timeline. A result whose shape really differs (say, one that followed a square start frame) is left as it came.

Editing keeps the source's aspect ratio and length — the model requires both — so those are not yours to set. Resolution is: it defaults to the source's, and raising it costs more, which the Generate button shows before you commit. The source has to be between 4 and 30 seconds; shorter or longer clips are refused before any credits are spent.

You can attach reference images to an edit — "make the jacket look like this." Connect an image node on the canvas or add one in the panel, then name it in the instruction as @Image1, @Image2 and so on; the numbers follow the order shown in the panel. Say nothing about them and the model still gets them as general guidance for the change.

Extending works in both directions. Adding to the end continues the clip and is billed on the added seconds only. Adding to the start builds the moments leading up to the shot; that route re-renders the whole result, so it is billed on the full length and the source plus the added time has to stay under 30 seconds. The Generate button always shows which one you are about to pay for.

When credits are charged

Credits are deducted when the generation is queued, not when it finishes. If the generation fails, the credits are refunded automatically — you do not need to ask. Failure messages say so explicitly.

Things that are free:

• Opening the recipe of an existing take to change it.
• Switching which take is the "best" one for a shot.
• Editing, trimming and rearranging the timeline.
• Rewriting a blocked prompt with the built-in suggestion.

Takes

A generation never overwrites anything. Each run adds a take to the node, and you pick which one is current. Old takes keep their own prompt, model and settings, so you can compare and go back.

This is why regenerating is cheap in attention even when it costs credits: the version you liked is still there.

When a generation fails

Failures fall into a few groups, and the message tells you which:

• Content policy. The model refused the prompt or a reference. Recognisable people, characters and brands are the usual cause.
• Input rejected. A reference file is the wrong shape, format or length for that model.
• Provider trouble. The model's provider is having a bad minute. Nothing to fix on your side; try again shortly.

In every case the credits come back. FAQ & troubleshooting goes through the common ones with what to change.

Running out of credits

If a generation would cost more than you have, it is blocked before anything is charged and you are offered the top-up or plan page. Nothing is queued and nothing is deducted.

Last updated 2026-09-15. Prices, model specs and credit costs on this page are read from the live catalogue at build time.