Pick the model. Or let the numbers pick.

Pick this when you need a real-looking person saying real words, or a still brought to life, at a normal budget.

Verified

• text: 2026-09-06
• image: 2026-09-06
• reference: 2026-09-06

Latency

about 3 minutes for a 5 s clip at 720p

Best for

talking-head UGC · product in hands · crazy look · bulk daily posts · animating a still (first frame) into a clip

Avoid for

clips over 15s · hero shots where 2.5 detail is worth 2x the price · anything that needs a seed

Modes

The fields you pass pick the mode. Credits are per output second at the chosen quality; the default is 720p.

• textNo seed. Every render is new; keep a series consistent with references, not seeds.
• imagefirst_frame becomes frame one of the clip; last_frame (optional) becomes the final frame and the model animates between them.
• imageFrames only: refs, video_refs and audio_refs are not accepted in this mode. To keep an identity AND set the frame, put the frame image in refs and describe it as @image1.
• imageaspect "adaptive" (default) follows the first frame. Frame images: jpeg/png/webp, ratio between 0.4 and 2.5, 300 to 6000 px, up to 30 MB.
• referenceAddress references in the prompt: "@image1 holds @image2 and says ...". Numbering is per list: refs are @image1.., video_refs are @video1.., audio_refs are @audio1...
• referenceReference clips: 2 to 15 s each, 15 s in total, mp4/mov 480p to 1080p, 24 to 60 fps, up to 50 MB. Their seconds are billed like output seconds.
• referenceAudio references: wav/mp3, 2 to 15 s each, 15 s in total, up to 15 MB. Audio alone is not accepted: add an image or video reference.
• referenceaspect "adaptive" follows the first video reference, else the first image, else the prompt.

Prompting

• 01Write the shot as a director: who (age, look), where (setting, light), what happens, phone framing, and the exact spoken words in quotes. About 2.3 words per second.
• 02Keep a person consistent across clips with refs (a portrait or a character sheet), addressed as @image1; the model has no seed.
• 03To animate a specific still, pass it as first_frame; add last_frame to control where the motion ends.

Last 30 days, live

From GET /v1/models, refreshed every few minutes. Auto score: an auto-judge grades every job 0 to 1.

Pick this when the user asked for the best possible single clip and accepts the wait and about 2x the price.

Verified

• text: 2026-09-05
• image: 2026-09-06
• reference: 2026-09-06

Latency

12 to 25 minutes per clip; plan the wait

Best for

hero product ads · close-up faces · one clip that has to be the best

Avoid for

drafts · bulk · anything where 2.0 is good enough: it is about 2x the credits · anyone who cannot wait 15 to 30 minutes

Modes

The fields you pass pick the mode. Credits are per output second at the chosen quality; the default is 720p.

• textNo seed.
• imagefirst_frame becomes frame one; last_frame (optional) becomes the final frame.
• imageaspect must be "adaptive" on this model in image mode: the clip takes the frame's ratio. Any other aspect is refused.
• imageFrames only: refs, video_refs and audio_refs are not accepted in this mode.
• referenceAddress references as @image1.., @video1.., @audio1.. (numbered per list).
• referenceReference clips: 2 to 30 s each, 30 s in total, up to 4K, up to 200 MB; their seconds are billed like output seconds. Audio references may stand alone here.
• referenceNever write edit or extend intent ("edit the video", "add", "remove", "replace", "extend", "continue") into a reference prompt: the provider reclassifies the task and fails it minutes later. Describe the NEW clip you want.

Prompting

• 01Same director-style prompt and @image1 references as seedance-2.0.
• 02In image mode leave aspect out (it is adaptive); the frame decides the ratio.
• 03Do not put edit or extend wording in a reference prompt.

Last 30 days, live

From GET /v1/models, refreshed every few minutes. Auto score: an auto-judge grades every job 0 to 1.

Pick this when you are making the image a video will be built on: a portrait, a character sheet, a first frame, or an edit that has to keep the same person.

Limits

1024x1024, 1024x1536, 1536x1024 · up to 4 refs

Verified

no recorded run yet

Latency

about 30 seconds, a little longer for an edit with references

Best for

portraits and character sheets that a video has to keep · the first frame of a clip · product in hand · edits that must not lose the face

Avoid for

bulk throwaway drafts where gpt-image-2.5-flare is faster

Prompting

• 01Concrete subject, age, framing, light, what the hands do; it follows layout instructions like "four poses on a plain background" or "headroom for a caption".
• 02With refs it edits or composes from them and holds the identity across poses, which is what makes a character sheet usable as a video reference.
• 03Pass the result straight to generate_video as first_frame (to animate it) or in refs (to keep that person across clips).

Last 30 days, live

From GET /v1/models, refreshed every few minutes. Auto score: an auto-judge grades every job 0 to 1.

Pick this when you want the same look as gpt-image-2.5 but faster, or you are making several images at once.

Limits

1024x1024, 1024x1536, 1536x1024 · up to 4 refs

Verified

no recorded run yet

Latency

about 15 to 30 seconds

Best for

variants and drafts at the same quality tier · batches of frames · anything where a few seconds matter

Avoid for

the one sheet a whole series depends on, where gpt-image-2.5 edits hold identity a little better

Prompting

• 01Same prompts as gpt-image-2.5; it is the speed tier of the same family.

Last 30 days, live

No runs in the last 30 days.

From GET /v1/models, refreshed every few minutes. Auto score: an auto-judge grades every job 0 to 1.

Pick this when you are reproducing something that was made on gpt-image-2; otherwise take the default.

Limits

1024x1024, 1024x1536, 1536x1024 · up to 4 refs

Verified

2026-09-05. every video pipeline stage A-C; portrait + sheet produced on run 2749ee84; standalone via generate_image

Latency

under a minute

Best for

the previous generation, kept selectable for runs that were built on it

Avoid for

new work: gpt-image-2.5 is the default and holds identity better

Prompting

• 01Concrete subject, age, framing, light, what the hands do; it follows layout instructions like "headroom for a caption".
• 02With refs it EDITS or composes from them (a product into a hand, a portrait re-lit); without refs it paints from the prompt alone.

Last 30 days, live

From GET /v1/models, refreshed every few minutes. Auto score: an auto-judge grades every job 0 to 1.

Pick this when you need a clean voice track and no face.

Limits

–

Verified

2026-09-05. wired in media-worker-v2 (tts.js, dubbing.js); standalone via generate_audio

Latency

seconds

Best for

voiceover on b-roll · narration · a standalone voice file · an audio reference for generate_video

Avoid for

lip-synced talking head: generate_video renders speech natively

Prompting

• 01Emotion tags like [excited] or [whispers] are honoured; keep sentences short for pacing.
• 02Seven named voices (sarah default) or a raw ElevenLabs voice id.

Last 30 days, live

From GET /v1/models, refreshed every few minutes. Auto score: an auto-judge grades every job 0 to 1.