> ## Documentation Index
> Fetch the complete documentation index at: https://docs.adskull.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Changelog

> New features, updates, and fixes shipped to the AdSkull API.

<Update label="August 11, 2026">
  ## Updates

  **MiniMax H3 (Hailuo 03).** [`minimax-h3`](/public/models#video-models) is an omni-modal video model with **native stereo audio always on** — no audio flag. Resolutions `768p` and `2K`; durations any integer `"4"`–`"15"`; ratios `21:9`, `16:9`, `4:3`, `1:1`, `3:4`, `9:16`. Input modes: text (ratio required) or first-frame image (`input_image_url` — the output matches the image, so omit `ratio`). From **35 credits** (768p, 4s).

  **MiniMax H3 Video-to-Video.** [`minimax-h3-v2v`](/public/models#video-models) runs the same model in `mode=edit` against a reference `input_video_url` (MP4/MOV, H.264/H.265, ≤ 50 MB, 2–15s per clip, ≤ 15s total), with `adaptive` as the default ratio. Billing is per second of `input + output` — 8 credits/s at 768p, 13 credits/s at 2K, rounded up to the nearest 10; `/v1/pricing` lists the 15s-reference ceiling and shorter references are charged on their measured duration.
</Update>

<Update label="August 8, 2026">
  ## Updates

  **Seedance 2.5 — 30 seconds in a single pass.** [`seedance-2.5`](/public/models#video-models) (ByteDance) generates up to 30 continuous seconds in one generation, no stitching. Durations `"4"`, `"5"`, `"8"`, `"10"`, `"15"`, `"20"`, `"30"`; ratios `1:1`, `4:3`, `3:4`, `16:9`, `9:16`, `21:9`, `adaptive`; `audio_enabled` is a free toggle. Input modes: text, first frame, first + last frame, or up to 4 reference images. From **50 credits** (480p, 4s).

  **Seedance 2.5 V2V.** [`seedance-2.5-v2v`](/public/models#video-models) runs the same model in `mode=edit` against a source `input_video_url`, with the same ratios, resolutions and durations. Credits are higher than the generate model because reference-driven generation is billed on the reference video's length as well as the output's, and the reference may be up to 30 seconds.

  <Note>
    Both Seedance 2.5 ids cap at **720p** — this endpoint exposes `480p` and `720p` only. There is no 1080p or 4K option (native 4K belongs to Seedance 2.0 coverage, not 2.5).
  </Note>
</Update>

<Update label="July 19, 2026">
  ## Updates

  **Five new image models.** The image catalog has grown from 10 to 15 models: [`nano-banana-2-lite`](/public/models#image-models) (budget Nano Banana, 3 credits, fixed 1K), `seedream-5-lite` (cost-efficient 2K/4K, from 3 credits), `gpt-image-2` (OpenAI next-gen, 1K/2K/4K), `gpt-image-1-5` (OpenAI flagship, medium/high), and `z-image` (photoreal drafts for 1 credit).

  **Four new video models.** [`grok-imagine-1.5`](/public/models#video-models) (xAI image-to-video, from 7 credits), `happyhorse-1.1` (Alibaba text/image/reference-to-video with 4:5 and 21:9 support), `kling-3.0-turbo` (faster Kling 3.0), and `gemini-omni` (Google multimodal, up to 4K).

  **Per-model aspect ratios.** `POST /v1/image-generations` now accepts every ratio a model advertises in `GET /v1/models/{model_id}` — including `4:5`, `5:4`, `2:3`, `3:2`, and banner strips like `8:1` — instead of a fixed six-value list.

  ## Bug fixes

  * `ratio: "4:5"` (and other model-supported ratios) no longer return `400 invalid_request`; ratio errors now list the accepted values for that model.
  * Omitting `quality` no longer fails for models without a `default` price tier — the model's default quality is applied automatically.
  * Image-to-video models (`wan-2.6`, `hailuo-2.3`, `grok-imagine-1.5`) and edit/motion models now fail fast with `400 invalid_request` when the required `input_image_url` / `input_video_url` is missing, instead of reserving credits and failing in the worker.
</Update>

<Update label="May 22, 2026">
  ## Updates

  **More LLMs available through a single endpoint.** The catalog has grown from 7 to 12 models. Call any of them through [`POST /v1/llm/chat`](/public/feature-llm-chat) — switching models is still a one-line change.

  **Simpler multimodal flow.** Vision, audio, and video inputs now go through the standard chat endpoint using `attachments`. Responses come back synchronously, so there's no job to poll. See [LLM Multimodal](/public/feature-llm-multimodal) for examples.

  ## Bug fixes

  * Corrected the model count in the LLM Chat overview (now reflects all 12 models).
</Update>