Skip to main content
Product updates for the Deepshi API. Subscribe with the RSS button to get new models and features as they ship. Model ids, context windows, and pricing are always current in GET /v1/models and on the models pages.
ModelsTextImageVideoMusic

New text, image, video, and music models

Added six image models (kling-image-o3, kling-image-v3, reve-2.1, krea-2-turbo, ernie-image, cosmos-3-super), two video models (seedance-2.0-mini, veo3.1-lite), one music model (lyria3-pro), and two chat models (inkling, kimi-k3). Recently added models now appear first on the text, image, video, and music pages.
ModelsText

New text models

Added the latest frontier chat models to the catalog, callable with the same key and balance as every other model. Browse them on the text models page or list them with GET /v1/models.
ModelsImageVideoMusic

New image, video, and music models

Expanded the image, video, and music catalogs with additional models. See image models, video models, and music models for the current lineup and per-model parameters.
MusicEndpoint

Music generation API

Launched the asynchronous music endpoints: POST /v1/audio/generations to create a song from a style prompt and lyrics, GET /v1/audio/generations/{id} to poll it, and /content to fetch the finished track. See music generation.
VideoEndpoint

Video generation API

Launched the asynchronous video endpoints: POST /v1/videos to create a job from a prompt or an image, GET /v1/videos/{id} to poll it, /content to redirect to the finished file, and DELETE /v1/videos/{id} to cancel. Jobs are billed once on completion; a cancelled or failed job is never charged. See video generation.
ImageEndpoint

Image generation and editing

Added POST /v1/images/generations and POST /v1/images/edits, with per-model sizing and quality parameters discoverable through GET /v1/models. Every image response returns the actual usage.cost. See image generation and image editing.
TextReasoning

Reasoning controls

Control reasoning depth with the reasoning_effort parameter (none/low/medium/high). The trace is returned as message.reasoning and message.reasoning_details. See reasoning.