Novita AI has issued a model API retirement notice covering a broad group of image, video and speech endpoints. The affected models are scheduled to stop on 19 August 2026, giving developers just over two weeks from the 3 August notice to migrate or stop sending new requests.
The change is larger than a single version replacement. It touches FLUX and Z Image generation, image utility functions, Wan and Seedance video, and MiniMax text-to-speech endpoints. Novita provides a successor for most entries, but not for every capability.
Image generation consolidates around newer endpoints
FLUX.1 Kontext Dev, Pro and Max are all listed for retirement, with Qwen-Image named as the text-to-image replacement. Novita gives the same Qwen-Image destination for the Flux 2 Dev, Flex and Pro Image Gen endpoints.
Z Image Turbo and its LoRA variant also move to Qwen-Image. These are not guaranteed drop-in changes: prompts, reference-image handling, output dimensions, latency and safety behaviour may differ between model families. The notice provides the target names but not a compatibility table or automated migration tool.
Three image utilities have no replacement listed: Image Eraser, Image Upscaling and Image Remove Background. Applications that depend on those endpoints will need another implementation rather than a model-ID change. That may mean selecting a different Novita API if one becomes available, moving the function to another provider, or running a suitable model separately.
Video users are directed to Wan 2.7
Wan 2.1 and Wan 2.2 text-to-video endpoints are scheduled to retire in favour of Wan 2.7 Text-to-Video. Their image-to-video counterparts similarly move to Wan 2.7 Image-to-Video.
Seedance 1.5 Pro is also affected. Both its text-to-video and image-to-video endpoints list the corresponding Wan 2.7 service as the replacement. That is a cross-family migration, so teams should not assume that an existing prompt, seed, duration or camera instruction will reproduce the same result.
For media pipelines, visual quality is only one test. Output aspect ratios, accepted input types, generation time, moderation, audio handling and cost can affect whether the replacement fits an automated workflow. Novita’s notice does not compare those characteristics.
MiniMax speech moves from 2.6 to 2.8
The speech changes are more direct. MiniMax Speech-2.6-hd moves to MiniMax Speech 2.8 HD, with separate synchronous and asynchronous long-form destinations. MiniMax Speech-2.6-turbo likewise moves to the 2.8 Turbo sync and async endpoints.
Even a versioned successor warrants regression testing. Voice identity, pronunciation, pacing, supported languages, long-form segmentation and response payloads can change. Applications that cache model identifiers or validate responses against a fixed schema should check both configuration and code paths before the deadline.
The short window makes inventory the first job
Novita’s required action is explicit: migrate where a replacement is listed or stop sending new requests before retirement. The notice does not state that traffic will be automatically redirected, and relying on an implicit fallback would risk production errors on 19 August.
A useful inventory should include the exact endpoint ID, whether calls are synchronous or asynchronous, expected inputs and outputs, volume, fallback behaviour and the owner of each consuming application. Tests should compare representative production prompts rather than a single demonstration request, particularly for video and image models where outputs are non-deterministic.
The announcement does not include pricing changes, regional differences, rate limits for the successors or an extension policy. Those details may determine whether a nominal replacement is operationally acceptable. The firm date, however, is clear: affected integrations need an intentional decision before 19 August, and the three image utilities without named replacements require the most substantial rework.