Skip to main content
v1.0.10
POST /v1/images, POST /v1/videos and POST /v1/audio now refuse a request without model with validation_error (422) before anything is saved or charged. These requests used to be accepted and then fail; a music request without one ran as lyria-002, so name that model to keep the old behaviour. A video draft upgrade (draft_generation_id) may still omit model. The estimate’s pricing_params description now says what it always did: a value that has its own argument (tier, generate_audio and the rest) is refused there, not overridden.
v1.0.9
eleven_v3 on POST /v1/audio now refuses previous_text, next_text, previous_request_ids and next_request_ids with validation_error (422) before anything is charged. ElevenLabs refuses them on that model, so these requests used to fail after submission. eleven_v4, eleven_multilingual_v2 and eleven_turbo_v2_5 still take them.
v1.0.8
New model depth-anything-video-fal (Video Depth Anything) on POST /v1/videos: one source_video returns a depth map of every frame. Five new optional fields set it up: depth_model (VDA-Small, VDA-Base, VDA-Large), depth_colormap (grayscale, turbo, inferno, magma, viridis), max_frames (1-2,400), output_fps (1-120; every frame is kept, so it changes playback speed) and side_by_side. tier picks the output resolution (auto, 360p-1080p). You are charged for the seconds processed, at most 2,400 frames.
v1.0.7
Model estimates include may_charge_more. When true, the displayed charge is an upfront estimate: actual provider usage may require additional credits after generation. Unused credits are refunded at the original discount. Availability follows each mode’s deployed Admin Catalog billing policy; explicit customer spend caps still apply.
v1.0.6
Model estimates accept video_input_durations, a JSON map of video roles to duration arrays in seconds (for example, {"reference_videos":[3.25]}). Estimates apply the selected mode’s configured input/output billing policy. Generation measures the actual files before charging; estimate durations do not determine the final reservation.
v1.0.5
seedance-2-5-fal adds an extend_video mode on POST /v1/videos: one source_video (4–30 s) and a chosen duration_seconds (4–30) return only the new part, starting where the clip ends or ending where it starts, as the prompt says. Send "mode": "extend_video": without it a lone source_video is still read as video_edit. Model pages now say where mode is required for this reason. Misreads return invalid_request.
v1.0.4
Adds request support for catalog-gated Seedance 2.5 drafts on POST /v1/videos: draft: true with tier: "480p" creates a preview; draft_generation_id selects a saved draft generation for a 1080p final. Only the creator’s completed, unexpired BytePlus direct drafts with saved task IDs and billing evidence are eligible. Finals reuse the original inputs and are billed independently. Availability follows the deployed Admin Catalog configuration.
v1.0.3
Two new speech models on POST /v1/audio: gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. Both take an optional delivery note (request-wide or per segments[] line, up to 200 characters) describing how a line is read, accept <laugh>, <sigh>, <short pause> and other angle-bracket vocal tags in the text, and accept any Extended Voice Library id as voice_name. delivery returns a 422 on models that do not declare it; GET /v1/models/{id} shows the parameter where it exists.
v1.0.2
GPT Image models now charge for their image inputs, as OpenAI does: each reference image and a mask add 1.32 credits, once per request. /v1/models/{id}/estimate takes mask_image_count (0 or 1). GPT Image 2 is now priced per exact size from its size list; a size outside that list, or auto, now returns a 422 instead of a flat band price, and a transparent background (which it cannot render) is refused before any charge. Like other models with input charges, an unrecognised mode on these models now returns a 422; omit it or send default.
v1.0.1
/v1/models/{id}/estimate now rejects a prompt_length above 1,000,000 with a 422. That is well above any model’s prompt limit, so real prompts are unaffected.
v1.0
Initial public release: organization API keys, /v1/models, /v1/models/{id} (with per-mode requirements and example_request), /v1/models/{id}/estimate, /v1/uploads (returns an asset_id), /v1/images, /v1/videos (text, image, reference, extension and motion control in one endpoint), /v1/audio, /v1/generations, /v1/generations/{id} (outputs carry asset_id), /v1/credits.Model ids containing slashes work in detail and estimate URLs. Model pricing includes the best discount applicable to your organization. Uploads accept up to 20 MiB for images, 1 GiB for video and 100 MiB for audio. Public uploads require the exact declared byte count so the stored file matches the asset’s recorded size and reserved storage quota.