GET /v1/models always returns the current ones, and POST /v1/cost quotes a request before you send it.
Kling Image 4.0 and Kling Image 4.0 Long
POST /v1/images/kling-image-4.0: Kling’s newest image model, with realistic images, correctly spelled text and edits that combine up to 10 reference images (refer to them in the prompt as @image_1, @image_2 …).size2K(the default, about 4 MP) or4K(5440x3072 at 16:9);aspect_ratioisauto(the default: the model chooses) or one of1:1,16:9,9:16,4:3,3:4,3:2,2:3,21:9.num_images(1–4), each generated and billed independently. No seed or negative prompt.
POST /v1/images/kling-image-4.0-long: one tall image, 1024 px wide: a full webpage design (skill: "webpage") or an ecommerce product detail page made from 1–3 product photos (skill: "ecommerce", length short, medium or long). For a webpage, reference_images takes up to 2: the subject first, then a style reference. The page’s text follows the prompt’s language.Pricing. Kling Image 4.0: 2K 75, 4K 115 credits per image. Kling Image 4.0 Long: a webpage 675; a product page 290 (short), 580 (medium) or 965 (long). Both add 4 credits for each reference image, per generated image.MCP. Both are in list_models and generate_image (skill and length for the Long model).MAI Image 2.7
POST /v1/images/mai-image-2.7: Microsoft’s newest image model, with realistic photos, correctly spelled text in images, and precise edits from plain instructions that change what you ask and keep the rest. Text-to-image, plus editing and combining up to 5 input images.size1K(the default, about 1 MP) or2K(about 4 MP);aspect_ratioisauto(the default: an edit keeps its first input’s frame and comes back aligned with it; text-to-image is1:1) or one of1:1,4:3,3:4,3:2,2:3,16:9,9:16(on an edit, a ratio other than the input’s own re-composes the picture into that shape).background: "transparent"returns a cut-out (PNG, or WebP withoutput_format: "webp") from a text prompt. With input images the model draws the subject anew; it is not a background remover.num_images(1–4), each generated and billed independently. Prompts up to 8,000 characters. No seed or negative prompt.
1K 50, 2K 190 credits, plus 11 (1K) or 45 (2K) credits for each input image, per generated image.MCP. mai-image-2.7 is in list_models and generate_image.Mureka V9.5
POST /v1/audios/mureka-v9.5: Mureka V9.5, Mureka’s newest music model. The same inputs as mureka-v9: a song sings your lyrics word for word (up to 5,000 characters, with [verse] / [chorus] tags), with an optional style prompt (up to 1,024 characters) and vocal_gender; set instrumental: true for an instrumental made from the prompt. One MP3 per request.Pricing. 200 credits per song or instrumental, a flat price whatever its length. mureka-v9 stays at 60.MCP. generate_music takes model: "mureka-v9.5".SeeDance 2.5 at 4K
POST /v1/videos/seedance-2.5 takes resolution: "4k": 3840 × 2160 at 16:9, 2160 × 3840 at 9:16 and 2880 × 2880 square, with every mode (text, frames, references, editing and extension) and any length from 4 to 30 seconds.- 4K output is H.265 (HEVC), 10-bit, like
1080p. See video output format. - Draft mode stays 480p, with its final at 1080p.
POST /v1/cost quotes it.MCP. generate_video takes resolution: "4k" on seedance-2.5.SeeDance 2.5 and 2.5 Flash: aspect_ratio with reference videos
On seedance-2.5 and seedance-2.5-flash, an aspect_ratio you send with reference_videos now sets the output’s shape. Leave it out and the model picks one (adaptive), as before. A job the model reads as an edit or an extension of a reference clip keeps that clip’s shape, and a start frame (image) still sets the shape on its own.SeeDance 2.5 Flash
POST /v1/videos/seedance-2.5-flash: SeeDance 2.5 Flash, ByteDance’s faster, lower-cost SeeDance 2.5. Text-to-video, start and end frames, and reference-to-video from up to 10 images, 5 video clips and 5 audio clips, which also edits or extends a clip. 480p or 720p with generated sound, 4–30 seconds.- Pricing: 93 credits per second of output at 480p, 200 at 720p. Reference videos add 57 (480p) or 123 (720p) credits per second of input.
POST /v1/costquotes any request. - For 1080p or draft mode, use
seedance-2.5.
seedance-2.5-flash is in list_models and generate_video.Kling 4.0 Flash, new Kling 3.0 prices, two Kling models retired
POST /v1/videos/kling-v4-flash: Kling 4.0 Flash, the faster, lower-cost Kling 4.0. 720p with generated sound, 3–20 seconds, from text, a first frame, up to 10 reference images, or a video to edit.- Pricing: 123 credits per second of output. With an input video, 82 credits per billed second: twice the output when the input is shorter than the output, otherwise input plus output.
- Kling 3.0 prices changed: Standard 219 credits per second (328 with audio), Pro 292 (437), 4K 546. Kling 3.0 Omni is 219 at 720p, 292 at 1080p and 546 at 4K (292 and 364 with audio at 720p and 1080p). An Omni edit bills each second of the input clip (3–10 s): 219 at 720p, 292 at 1080p. Kling 3.0 and Omni take 3–15 seconds and prompts of up to 2,500 characters.
POST /v1/costquotes any request exactly. - Retired:
kling-v2.6-proandkling-3.0-turboanswer410 model_retired. Usekling-4.0andkling-v4-flashinstead.
kling-v4-flash is in list_models and generate_video.Eleven Music v2.5
POST /v1/audios/eleven-music-v2.5: ElevenLabs’ newest music model. Songs with vocals in many languages, or instrumentals, from a text prompt; lyrics are sung as written.prompt(required, up to 2,000 characters),lyrics(optional, up to 3,000; the two together up to 4,090),instrumental(not withlyrics),duration10–300 seconds (default 60). MP3 output.- Prompts that name real artists, songwriters, songs, albums or record labels, or quote a real song’s lyrics, are refused with a suggested rewrite. Refused jobs are not billed.
duration, billed per second with a 10-second minimum: 30 s = 100, 60 s = 200, 180 s = 600 credits. GET /v1/models returns credits_per_unit: 200 with unit_type: minute, and POST /v1/cost quotes the exact price.MCP. eleven-music-v2.5 is in list_models and generate_music, with duration, lyrics and instrumental.MiniMax H3 Image
POST /v1/images/minimax-h3-image: MiniMax’s image model, with exact text on signs, menus and posters, natural light and fine detail. Text-to-image, plus editing and multi-reference composition from up to 10 reference images.size2K(the default) or4K;aspect_ratioisauto(the default: an edit keeps its first reference’s shape) or one of 13 shapes from9:21to21:9.num_images(1–4) are made in one call; an image the model’s safety check withholds is not billed.- No seed, negative prompt or transparent background.
2K 34, 4K 100 credits, plus 12 credits for each reference image, per generated image.MCP. minimax-h3-image is in list_models and generate_image.Seedream 5.5
POST /v1/images/seedream-5.5: ByteDance’s newest Seedream, with more natural light, materials and skin, sharper small and dense text, and edits that change what you ask and keep the rest of the picture as it was. It generates from text and edits or combines up to 10 reference_images.background: "transparent"returns a cut-out PNG from a text prompt alone or from references. See transparent backgrounds.size1Kor2K(the default);aspect_ratio: "match_input_image"keeps an edit’s shape.output_formattakespng(the default) orjpeg.- No grouped generation:
num_images(1–4) makes each image separately.
1K 26, 2K 46 credits, plus 10 credits for each reference image, per generated image.MCP. seedream-5.5 is now the default model of generate_image (it was seedream-5.0-pro), background: "transparent" included.GPT Image 2.6 Flare and Sunburst
POST /v1/images/gpt-image-2.6-flare and POST /v1/images/gpt-image-2.6-sunburst: OpenAI’s GPT Image 2.6. Flare is the fast one; Sunburst is for dense text, complex layouts and precise edits. Both generate from text and edit or combine up to 16 reference_images.complexity(low,mediumorhigh, the default) sets how much the model plans before it renders. It does not change the price.background: "transparent"returns a cut-out with a transparent background, from text or an edit, as a PNG (or a WebP). See transparent backgrounds.- Sizes: 1:1 at
1Kor2K, 16:9 and 9:16 at2Kor4K(3840 × 2160), the other ratios at one size.aspect_ratio: "auto"lets the model choose, and an edit keeps its picture’s shape.output_formatalso takeswebp.
quality, the same at every size and with any number of references: low 70, medium 120 (the default), high 175, xhigh 340, max 675 credits.Catalog. Image models priced by quality now list each tier’s price as quality_pricing in the model catalog the MCP server reads, and the MCP’s list_models shows it.MCP. generate_image takes both models, background: "transparent" included.SeeDance 2.5 draft mode: a 480p draft first, then its 1080p final
POST /v1/videos/seedance-2.5 can make a cheap draft and, once you like it, render the same video at 1080p: the same motion, framing and sound, sharper. See draft mode.draft: truemakes a 480p draft.GET /v1/videos/{id}on it returnsdraft_finalizable,draft_expires_atanddraft_final_credits.from_draft_job_id, the draft’sidsent on its own, renders its 1080p final within 7 days. The final reuses the draft’s prompt, inputs, length, aspect ratio, seed and audio setting, and its job returnsfrom_draft_job_id. Drafts made in the Video Workshop or over MCP work too.promptis no longer required whenfrom_draft_job_idis sent.
POST /v1/cost quotes both; send from_draft_job_id to price a final.MCP. generate_video takes draft on seedance-2.5, and the new finalize_video_draft tool makes a draft’s 1080p final from its job_id. check_generation_status shows a finished draft’s final price and deadline.Wardrobe: dress a character in a chosen outfit
Write an outfit’s@handle next to a character in any image or video prompt and the character wears exactly those clothes: "@barry wearing @camel_belted_overcoat_outfit on a rainy street". Up to 4 outfits per generation, beside up to 4 characters; their pictures count as reference images. A free library of outfits comes with every account, and you can make your own from a photo or a description on the web. See the Wardrobe guide.GET /v1/wardrobelists the outfits you can use: yours, the ones you saved and the free library. Free.GET /v1/wardrobe/{handle}gives one in full, with its full-size picture. Free.
search_wardrobe and get_wardrobe_item find outfits; generate_image and generate_video take the same @mention.Kling 4.0
POST /v1/videos/kling-4.0: Kuaishou’s Kling 4.0, with generated sound, in one endpoint: text-to-video, a first and an optional last frame, up to 10 reference images, reference videos, and editing a video (the output keeps its length and shape). 720p or 1080p, 3 to 30 seconds, 24 fps. Name inputs in the prompt as @image_1, @video_1 and so on.Pricing. 207 credits per second at 720p and 310 at 1080p. With any input video, 124 / 186 per billed second, where the billed seconds are the output plus the longer of the input videos and the output: editing a 5-second clip bills 10 seconds.MCP. generate_video takes kling-4.0 with all of its inputs: frames and reference images combine, and video edits a clip.Meshy 7.1: pick the move your character performs
animation_moveonPOST /v1/3d/meshy-7.1picks the move an animated character performs: one of about 650 fromGET /v1/3d/moves(dances, waves, cheers, walks, stunts and fights). Free; without it the move is 92, Double Combo Attack, as before.- Rigged jobs now also return a walk and a run:
walking_glb,walking_fbx,running_glbandrunning_fbxinfiles.
generate_3d_model takes animation_move on meshy-7.1 (it turns rigging and animation on by itself).More reference inputs for SeeDance; Wan 3.0 bills reference video
- SeeDance 2.0, 2.0 Fast and 2.0 Mini take
reference_videos(up to 3, 15s combined, billed per input second at half the output rate) andreference_audios(up to 3 withaudio, 15s combined, free) next to up to 9reference_images. With avideoto edit, up to 2 morereference_videoscan join it (thevideois “Video 1”), and edits take up to 9 reference images (was 6). - SeeDance 2.5 takes up to 30
reference_images, 10reference_videosand 10reference_audios(was 15 / 5 / 5), and a clip may run up to 30 seconds (still 30s combined per kind). - Wan 3.0 and Wan 3.0 Prime now bill reference-video input per second at the output rate, as they bill output: a 5-second 720p edit guided by an 8.2s clip costs
(5 + 9) × 120 = 1680credits (was 600). Reference images and audio stay free.POST /v1/costquotes it exactly.
generate_video takes reference_videos and reference_audios on seedance-2.0, seedance-2.0-mini, wan-3.0-video and wan-3.0-video-prime, with the caps above.Characters: cast AI influencers with an @mention
Write@handle in an image or video prompt to cast a character: one of the free BudgetPixel AI influencers, or a character of your own. Its picture and look go with the mention, so the same face appears in every post.- @mentions in the
promptof every image and video model that takes characters (each such model’spromptsays so).@handle.tagcasts a specific picture; up to 4 characters per generation. Create responses list the cast incharacters, andPOST /v1/costcounts them. GET /v1/characters: the characters you can cast, influencers first, withqandsourcefilters.GET /v1/characters/{handle}: one character in full, with the look word for word, the profile, directing notes and every picture with a download link (full size for the influencers and your own characters).POST /v1/characters/recommend: the influencers who fit a brief, each with why, an idea and a ready-to-use prompt.- The influencers are never shown nude or in sexual content: such requests are refused with
restriction_reason: "official_character_nudity". Suggestive outfits and poses are fine.
list_characters, get_character and recommend_influencers; generate_image and generate_video cast characters from @mentions in the prompt.Guide: AI influencers and characters.Job timing, prompt, output expiry and Retry-After on every 429
Job status responses now say when a job started and finished, return its prompt and say when its files expire, and every429 carries Retry-After.GET /v1/images/{job_id}andGET /v1/audios/{job_id}: newprompt,started_atandcompleted_at.started_atminuscreated_atis the time the job waited in the queue. Video jobs already returned all three.- Image, video and audio job status: new
expires_at, 24 hours aftercreated_at, andexpired. Afterexpires_atthe files are deleted and the response returns no URLs, where it used to return URLs that answered403. Each signed URL is valid for at least 6 hours; call the status endpoint again for fresh ones. GET /v1/videos/{job_id}now returnsid, like the create response and the API reference.job_idstays, with the same value.- Every
429sendsRetry-After. A full concurrency queue sends a suggested wait (5 seconds for images, 10 for audio, 15 for video) and the same body for images, video and audio:code: "queue_full",queue_limit,active_jobsandretry_after_seconds. The monthly spend limit sends the seconds until the month resets, and the posting limits send the time until you can post again. - A full queue is now refused before your inputs are fetched or checked, so the
429comes back at once instead of after a few seconds. Simultaneous requests can no longer start more jobs than your concurrency cap. - When your monthly spend limit is reached, job status,
GET /v1/models,GET /v1/account/creditsandPOST /v1/costkeep working, so you can still collect jobs that started before the limit.
Depth maps: face mesh
Depth maps can now carry each face’s mesh, so a video model given the depth map keeps lip, eye and head movement. Faces are found automatically.POST /v1/depth-maps/video-depth-anything: newface_mesh(draws the mesh over each face),max_faces(1 to 4, default 4) andinclude_face_landmarks(returns a JSON file of 478 landmarks, 52 expression scores and head pose per face per frame).GET /v1/depth-maps/{id}: newface_mesh(how much of the clip had a face, and each face’s frames) andface_landmarks_url.
size.MCP. generate_depth_map takes face_mesh, max_faces and include_face_landmarks; check_generation_status returns the face coverage and the landmarks link.Guide: Face mesh.Depth maps: Video Depth Anything
Turn a video into a depth map video: in the default grayscale, near is white and far is black, and the result is steady from frame to frame. Use it as a structure or motion reference for video models.POST /v1/depth-maps/video-depth-anything: avideo(a public URL or an uploaded-file URL; up to 30 seconds, MP4, MOV or WebM, up to 100 MB). PollGET /v1/depth-maps/{id}for the result. Job ids start withdpt_.size:small(the default, fastest),base(adds edge detail) orlarge(the finest detail and the steadiest).colormap:grayscale(the default),turbo,inferno,magmaorviridis.resolution:auto(the default),360p,480p,720por1080p.keep_audioandside_by_side(the input on the left, the depth map on the right) are off by default.- The result is an H.264 MP4 at the input’s frame size (up to 1080p) and frame rate (up to 30 fps).
video_urlis a short-lived signed URL, and the file is deleted 24 hours after generation. GET /v1/modelslists the model withtype: "video-depth"and a newsize_pricingfield, a map of eachsizeto its credits per second:{ "small": 20, "base": 30, "large": 40 }.POST /v1/costreturnscredits_per_second,size_pricingandmin_seconds, plus the exactcreditswhen you sendseconds.
small, 30 for base and 40 for large. The price is the same at every resolution.MCP. New tool: generate_depth_map.Guide: Depth maps API.OmniHuman 2.0
The newest version of ByteDance’s OmniHuman. It animates the person, character or pet in an image to speak your audio, with synchronized lips, expression and body motion, and camera movement you can direct with a prompt. OmniHuman 1.5 stays available.POST /v1/lip-sync/omni-human-2.0: animageand anaudioclip, plus an optionalpromptfor the scene, gestures and camera (Chinese, English, Japanese, Korean, Spanish or Indonesian). PollGET /v1/lip-sync/{id}for the video.- 720p. Audio can be MP3, WAV or OGG, up to 120 seconds and 15 MB, twice the length OmniHuman 1.5 accepts. Images can be 300 to 6000 px a side, with an aspect ratio between 2:5 and 5:2.
- Lip-sync jobs on every model now reserve their quoted price when they’re created, and the charge on success equals the quote from
POST /v1/cost.
FLUX 3 Video
Black Forest Labs’ video model, with generated audio: sound effects, ambience, and dialogue with lip sync. It can also extend a clip you supply.POST /v1/videos/flux-3-video: text-to-video; image-to-video from a first frame, or a first and a last frame; and extension of your own MP4 or MOV (1 second to 2 minutes, up to 100 MB). An extension returns your clip followed by the new footage.- 720p, 1080p, 1440p or 4K at 24 fps, 5 to 20 seconds. An extension adds 5 to 15 seconds.
generate_audio: falsemakes a silent clip at the same price.
FLUX 3 Klein
A fast, low-cost version of FLUX 3 from Black Forest Labs, from 11 credits per image.POST /v1/images/flux-3-image-klein: text-to-image and editing with up to 10 reference images (free), 1 to 4 images per request.- 768px, 1K, 1.5K, 2K or 4K, with
aspect_ratio: autoor one of 13 fixed ratios. It has no web grounding; use FLUX 3 for that.
FLUX 3 Video Upscale
A second video upscaler next to Topaz: 1.5x to 3x, in aprecise mode or a creative mode that adds detail.POST /v1/video-upscales/flux-3-video-upscale, then pollGET /v1/video-upscales/{id}.- Input up to 20 seconds, 2560x1440 and 100 MB.
upscale_factoris 1.5, 2, 2.5 or 3, with an optionalprompt. The frame rate and audio are kept.
precise mode and 267 in creative.3D models: Tripo H3.1 and Meshy 7.1
Turn one image into a textured 3D model you can download.POST /v1/3d/tripo-h3.1: the most faithful to the source image, cartoons included, as GLB with PBR materials. Options:texture,texture_quality(standardordetailed) anddetailed_geometry.POST /v1/3d/meshy-7.1: a lighter model (about 30K triangles) as GLB, FBX, USDZ and STL, with optionalriggingandanimationfor humanoid characters.- Poll
GET /v1/3d/{job_id}every 15 to 30 seconds; a model takes 2 to 5 minutes. Files are deleted after about 24 hours, so download them promptly. GET /v1/models?type=3dlists both models.
generate_3d_model.P-Video 2 Pro
A low-cost video model from Pruna AI, with generated audio and amode that trades cost against speed and quality.POST /v1/videos/p-video-2-pro: text-to-video, and image-to-video from a first frame, a last frame, or both. 480p or 768p, 5 to 15 seconds, 24 fps.
cost mode, 30 in speed and 60 in quality. At 768p: 37, 52 and 112.Stock sound effect and music search
Search BudgetPixel’s free sound effect and background music libraries, licensed CC BY 4.0. Like stock image search, it works with keys on the Free plan.GET /v1/stock/sound-effectsandGET /v1/stock/music: filter byquery,category(sound effects only),min_duration,max_durationandsort, up to 20 results per page.- Each result carries a ready-made credit line (
attribution.textandattribution.html) to include wherever you use it. - Stock image search also gained
transparent=true, for cut-outs with a transparent background.
search_stock_audio, and search_stock_images now takes transparent.Guides: Stock sound effects and Stock music.Nano Banana 2.1
Google’s image model, with sharper output and cleaner text in images than Nano Banana 2, and optional grounding in live web results.POST /v1/images/nano-banana-2.1: text-to-image and editing with up to 9 reference images (free), 1 to 4 images per request, at 1K, 2K or 4K.google_searchgrounds the image in live Google Search results;image_searchlets the model look at web images first. Both are off by default.
google_search adds 10 credits per image, and image_search adds 20 in total, with or without google_search.Licenses on posts
POST /v1/posts takes a license: arr (All Rights Reserved, the default), cc-by (CC BY 4.0) or cc0 (CC0 1.0). It can’t be changed after publishing.MCP. create_social_post takes license.FLUX 3
Black Forest Labs’ FLUX 3: text-to-image and editing with up to 10 reference images.POST /v1/images/flux-3-image: 1 to 4 images per request at 768px, 1K, 1.5K, 2K or 4K. Reference images are free.groundinglets the model research the prompt with web and image search before rendering. It’s on by default and costs nothing extra; set it tofalsefor faster, self-contained generation.GET /v1/modelsnow lists per-size prices for image models (resolution_pricing).
SeeDream 5.0 Flash
ByteDance’s fast, low-cost image model, typically done in under 20 seconds.POST /v1/images/seedream-5.0-flash: text-to-image and editing with up to 10 reference images (free), 1 to 4 images per request, at 1K or 2K.background: transparentkeeps a transparent background when you edit a single PNG cut-out, and returns a PNG with an alpha channel. SeeDream 5.0 Pro accepts it too.
MiniMax H3 Max
MiniMax’s H3 tuned for stronger prompt adherence and looks, with native audio, reference clips and extension.POST /v1/videos/minimax-h3-max: text-to-video; image-to-video from a start and/or end frame; reference-to-video with up to 5 images, 3 video clips and 3 audio clips; and extension of a video you supply.- 480p, 768p or 1080p (2K when extending), 5 to 15 seconds.
POST /v1/cost gives the exact figure.Stock image search
Search BudgetPixel’s free stock image library, licensed CC BY 4.0. Keys on the Free plan can call it; the rest of the API needs a paid plan.GET /v1/stock/images: filter byquery,category,orientation,color,typeandsort, up to 20 results per page. Each result carries a credit line to use with the image.
search_stock_images.SeeDance 1.5 Pro retired
SeeDance 1.5 Pro has been retired. Requests toPOST /v1/videos/seedance-1.5-pro now return 410 model_retired; use seedance-2.0 instead. See Retired models.