AI video model release tracker (2026): launch dates and access
Quick Answer
Quick answer
The most consequential 2026 AI video releases so far are Kling VIDEO 3.0 on February 5, Seedance 2.0 on February 12, Seedance 2.5 on July 31, MiniMax H3 on July 31 with open weights on August 3, and LTX-2.5 support on August 11. The largest access change is OpenAI's Sora shutdown: its web and app experience ended April 26, and the API is scheduled to end September 24. This tracker was verified from primary sources on September 13, 2026.
A release date does not prove broad availability, quality leadership or support inside every platform. The chronology below separates the official announcement from the action a creator or developer should take. Magic Hour's current catalog is checked separately against its model reference.
2026 AI video model release timeline
Release or change | Date | What changed | Current action |
|---|---|---|---|
October 15, 2025 | Google released Veo 3.1 and 3.1 Fast in paid Gemini API preview, with reference images, extension and first-to-last-frame transitions. | Use the current Gemini or platform model ID; verify preview or stable status before production. | |
February 5, 2026 | Kuaishou launched VIDEO 3.0 and VIDEO 3.0 Omni with multimodal input, native audio and clips up to 15 seconds. | Evaluate standard and Omni separately; do not treat the platform name as one fixed model. | |
February 12, 2026 | ByteDance launched a multimodal audio-video model with text, image, audio and video references and up to 15-second output. | Keep it in older workflow comparisons, but use 2.5 for a new latest-version test. | |
July 31, 2026 | ByteDance added up to 30-second audio-video generation, multi-round extension and stronger reference-based editing. | Re-test duration, extension and reference adherence instead of carrying forward 2.0 results. | |
July 31, 2026 | MiniMax launched H3 for unified text, image, video and audio context with native stereo audio and up to 15-second, 2K hosted output. | Separate hosted 2K results from the open base checkpoint when comparing quality or cost. | |
August 3, 2026 | MiniMax released H3 weights; its hosted Context-IR and Regenerate-2K stages remain separate services. | Record whether a result came from local H3-Base or the complete hosted workflow. | |
August 11, 2026 | Lightricks' official LTX-2 release added 2.5 checkpoint support; current docs list Fast and Pro variants up to 4K. | Record checkpoint, variant, resolution and local versus API execution. | |
April 26 / September 24, 2026 | OpenAI ended the Sora web and app experience April 26 and schedules the Sora API to end September 24. | Export existing work and migrate any remaining API dependency before shutdown. |
A red ceramic mug sits on a wooden café table beside a window. Steam rises slowly while the camera makes a gentle push-in. Soft morning light, realistic materials, one continuous shot, no people or readable text.
Test current AI video models in one workflow
Compare a short representative prompt across currently available models, then keep the model and settings with the best accepted-output cost for your workflow.
Try AI Video GeneratorFebruary 2026: Kling and Seedance move to multimodal creation
Kuaishou announced Kling VIDEO 3.0 on February 5 as part of the Kling 3.0 family. The release combines text-to-video, image-to-video, reference-to-video and in-video editing, with native audio and outputs up to 15 seconds. VIDEO 3.0 and VIDEO 3.0 Omni are distinct variants, so a comparison should name the exact one used.
ByteDance launched Seedance 2.0 on February 12. Its model accepts mixed text, image, audio and video inputs, supports reference-based generation and editing, and produces up to 15-second audio-video clips. Those facts make it a meaningful version boundary; they do not establish that every third-party interface exposes every control.
July 2026: longer Seedance output and MiniMax H3
Seedance 2.5 launched on July 31 with up to 30-second audio-video generation, multi-round extension and broader reference workflows. Any article still calling Seedance 2.0 the current version should be rechecked, while historical 2.0 tests should retain their original model label.
MiniMax announced H3 the same day. The provider describes unified text, image, video and audio context, native stereo audio and hosted output up to 15 seconds at 2K. On August 3, MiniMax separately released H3's model weights. The open H3-Base checkpoints and the complete hosted workflow are not equivalent: Context-IR and the Regenerate-2K stage remain hosted services.
August 2026: LTX-2.5 enters current workflows
The official LTX-2 repository added LTX-2.5 support in release v1.2.0 on August 11. The current LTX-2.5 documentation describes Fast and Pro variants, portrait and landscape output up to 4K, and text-to-video, image-to-video and audio-to-video endpoints.
For reproducibility, record whether the test used the full or distilled checkpoint, Fast or Pro, local weights or API, resolution, duration and audio setting. Saying only “LTX-2.5” can hide material differences in speed, output and cost.
The access change that matters most: Sora is shutting down
OpenAI's Sora discontinuation notice says the web and app experiences ended April 26, 2026, and the API will end September 24, 2026. This is an access event rather than a new model release, but it belongs in a production tracker because it creates a firm migration deadline.
Do not begin a new long-term Sora dependency. Export existing content, identify prompts and assets that can be retained lawfully, choose a replacement model, run a matched test, and move production before the API cutoff.
How to read an AI video release announcement
Name the exact version. A platform can expose several models and a model can appear in several platforms.
Separate announcement from access. Record the release date, rollout region, waitlist or plan, API status and the date you personally verified access.
Separate hosted and open workflows. Provider-side prompt enhancement, preprocessing or upscaling can make a hosted result differ from local weights.
Verify exposed controls. Duration, resolution, aspect ratio, audio, reference inputs and editing may vary by endpoint or integration.
Treat demos as demonstrations. A curated clip does not establish reliability, accepted-output cost or performance on your own work.
When should you switch models?
Switch only when a release addresses a measured constraint: a required input, longer usable duration, stronger reference adherence, acceptable audio, supported resolution, reliable API access or lower cost per accepted output. A newer version number alone is not enough.
Use a fixed prompt set and permitted reference assets. Keep the old and new model outputs, settings, generation time, spend and rejection reasons. Compare blind where practical, then calculate accepted-output cost as total generation spend plus review and correction time divided by approved clips.
Frequently asked questions
As of September 13, the dated releases tracked here include Kling VIDEO 3.0, Seedance 2.0 and 2.5, MiniMax H3 and LTX-2.5. “Newest” changes quickly, so verify the linked primary source and access status before citing it.
Kuaishou announced the Kling 3.0 model family on February 5, 2026.
ByteDance announced Seedance 2.5 on July 31, 2026, with up to 30-second generation and multi-round extension.
MiniMax released H3 model weights on August 3, 2026. Its complete hosted workflow includes Context-IR and Regenerate-2K services that are not part of the base-weight release.
The Sora web and app experience ended April 26, 2026. OpenAI says the Sora API will end September 24, 2026.
Related current guides
Use the best AI video generators guide for platform selection, the AI video model benchmark for a reproducible test method, and the state of AI video and image generation for the broader market context.
Runbo Li is the Co-founder and CEO of Magic Hour, where he builds AI video and image tools for content creation. He is a Y Combinator W24 founder and former Data Scientist at Meta, where he worked on 0-1 consumer social products in New Product Experimentation. He writes about AI video generation, AI image creation, creative workflows, and creator tools.
