After Sora: the AI video APIs still taking requests

OpenAI's deprecation list gives the Videos API and every sora-2 and sora-2-pro snapshot a removal date of 24 September 2026, and in the column for a recommended replacement it gives nothing at all. Anything built on Sora has to move to someone else's model, and the request shapes differ enough that this is a port, not a find-and-replace.

xAI: Grok Imagine

Priced the way Sora was, per second of output, with three video models to choose from. On xAI's models page as checked on 6 October, grok-imagine-video-1.5-lite costs $0.020 a second, grok-imagine-video $0.050 and grok-imagine-video-1.5 $0.080. Video 1.5 does text-to-video, image-to-video and reference-to-video, with native 1080p for the first two. One detail worth knowing before you port prompts: xAI's release notes say its text-to-video runs as a text-to-image step followed by image-to-video, so the first frame is a still the model chose.

Google: the Gemini API

Google's video models live in the Gemini API. The newest, gemini-omni-1.1-flash, became generally available on 27 August, with video extension, interpolation between two images, and output at 360p, 720p, 1080p or 4K. Beside it is veo-3.1, where Google's changelog sent users of veo-2.0 and veo-3.0 when it shut those down on 30 June — a reminder that a migration target can have a migration of its own.

MiniMax: H3 through the API

MiniMax serves H3, its model for video and sound generated together, through the API at platform.minimax.io. Its licence notes say the API is available globally, which matters because the open weights are not: the H3 licence excludes the European Union, the United Kingdom, South Korea and the United States. For a team in those places the hosted API is the way to use H3 without applying for a separate licence.

Kling

Kling sells API access through its developer platform. Kling 4.0 is in limited early access, with the full model due in October, so check the platform for which Kling models the API serves before you plan around 4.0's thirty-second clips.

What does not carry over

Prompts. Each of these models has its own idea of how a scene should be written: xAI and Kling take references and keyframes in their own ways, MiniMax H3 has a structured format with numbered shots and tagged dialogue, and Lightricks' own guide warns against pasting another model's prompt into LTX unchanged. Budget time for rewriting the prompts, not just the client code.