The 8 AI video models with the longest single generation, ranked

The pitch from every AI video maker this year is the same: a clip long enough to carry a scene. The claim is now worth checking against the makers' own numbers, because the gap between the longest single generation on this beat and the shortest is nearly four times. This list ranks the major AI video models by the longest single generation each maker says its own model can produce, longest first, as of 7 November 2026.

The rule for the ranking is fixed. A model is measured only by a figure the maker prints on a page this article opened. A comparison article, a reseller's listing or a reviewer's summary does not count, and where a maker's own documentation states no maximum, the model is left out rather than filled in. Pika is the case in point: its own site states no single-generation maximum for its own model, only a 30-second ceiling for the third-party models it hosts. Every figure below is one generation from a prompt or a still, the maker's own stated maximum. Clip limits move with every update, and a platform that hosts a model can cap it below what the model allows. Where a maker states a longer figure for a video extension or an edit pass, that figure is named in the entry.

Eight models survived the test. Google's Veo 3.1 preview was shut down in the Gemini API on October 22, 2026, but the model stays generally available through Google Cloud, and that documentation carries the limit used below. No hands-on tests were run for this list; it weighs the makers' own published generation limits and nothing else. Three models tie at the top at 30 seconds, and they are ordered here by how much they let a creator steer that one generation.

8. Veo 3.1 (Google)

Veo 3.1 generates video up to 8 seconds long in a single pass, the shortest maximum of any model here. Google Cloud's own model page for Veo 3.1 states video lengths of 4, 6 or 8 seconds, with the 8-second setting required when a reference image is used. The same page lists output at 720p, 1080p or 4K, 24 frames a second, native sound generation, and video extension as a supported feature, without stating a length for the extension.

Veo 3.1 is generally available through Google Cloud. In the Gemini API it was offered as a preview until Google shut that preview down on October 22, 2026, according to Google's Gemini API video documentation, which now directs developers to Gemini Omni Flash. Its 8-second ceiling is the shortest of the eight, and nearly four times under the three models tied at the top.

7. Hailuo 2.3 (MiniMax)

MiniMax's Hailuo 2.3 tops out at 10 seconds, and only at its lower resolution. MiniMax's own models page lists Hailuo 2.3, for text-to-video and image-to-video, at 1080p for 6 seconds and at 768p for 6 or 10 seconds, at 24 frames a second, so the 10-second generation is a 768p one. The page files Hailuo 2.3 under legacy models, behind MiniMax's newer H3 video model.

That resolution trade-off is the catch. A creator who wants Hailuo 2.3 at its sharpest 1080p gets 6 seconds, and the 10-second maximum only holds at 768p. The page lists no native audio for Hailuo 2.3. At 10 seconds, Hailuo 2.3 sits level with Luma's Ray3.2 and well below the three 30-second models at the top.

6. Luma Ray3.2 (Luma AI)

Luma's Ray3.2 generates 5 or 10-second clips from a prompt or a still, and goes longer only when it is transforming footage that already exists. Luma's own Ray page states that text-to-video and image-to-video produce 5 or 10-second native clips, that its Modify Video mode can produce up to 20 seconds depending on frame rate, and that Reframe takes source clips up to 16 seconds.

Ray3.2 output runs from 360p to 1080p, and the page says the text and image modes produce no native audio, with voiceover, music and sound effects added separately. On the single-generation measure used here, Ray3.2's maximum is 10 seconds, which places it below Gemini Omni 1.2 Flash and level with Hailuo 2.3. Its 20-second ceiling applies to reworking an existing clip, not to a fresh generation, so it does not move Ray3.2 up the list.

5. Gemini Omni 1.2 Flash (Google)

Google's Gemini Omni 1.2 Flash generates up to 15 seconds in a single generation, up from 10 in the model it replaces. Google's release notes dated November 5, 2026, announce the model as generally available with video generation of up to 15 seconds, first-and-last-frame interpolation of 3 to 10 seconds, and per-turn video extension of 3 to 10 seconds up to a 30-second total length.

The same entry lists character consistency across up to 6 reference images, a 30% lower video output price at $12.00 per 1M tokens, and doubled paid-tier rate limits. Google says requests to the previous Gemini Omni 1.1 Flash will route to 1.2 Flash from December 3, 2026. At 15 seconds, Gemini Omni 1.2 Flash is the longest model here that is not tied at the top, and it beats the older Veo 3.1 by nearly double.

4. LTX-2.5 (Lightricks)

Lightricks' LTX-2.5 reaches 20 seconds in one generation, the longest of any open-weights model on this list. Lightricks' own API documentation lists the fast variant, ltx-2-5-fast, with durations of 6, 8, 10, 12, 14, 16, 18 and 20 seconds at 720p and 1080p at 24 or 25 frames a second. The pro variant, and every tier at 1440p and 4K, stop at 10 seconds.

The model ships with open weights. Lightricks' model page says commercial use is free under $10M in annual revenue, and that a higher-revenue business needs a paid commercial agreement. An automatic-duration option lets the model set a clip's length from the prompt. At 20 seconds, LTX-2.5 is the longest of the open models and a clear step up from Gemini Omni 1.2 Flash, though still 10 seconds short of the three models tied at the top.

3. Wan 3.0 (Alibaba)

Alibaba's Wan 3.0 generates up to 30 seconds of video in a single pass, one of three models here tied at the top. Alibaba Cloud's Wan3.0 announcement states a maximum single-generation duration of 30 seconds, with smart duration recommendation and video extension, and lists output at 480P, 720P and 1080P.

The same page prices the API per second of generated video at $0.05 for 480P, $0.10 for 720P and $0.20 for 1080P, which puts a full 30-second 1080P generation at $6.00. Wan 3.0 also accepts documents, spreadsheets and PDFs as input, and Alibaba's page names audio texture and on-screen text as the areas still being improved. It ties with Seedance 2.5 and Kling 4.0 on the headline number, and it is the one of the three priced openly per second at every resolution.

2. Seedance 2.5 (ByteDance)

ByteDance's Seedance 2.5 generates 30-second audio-video clips in a single pass, the model's own launch post states. ByteDance says Seedance 2.5 extended single-pass generation from 15 to 30 seconds, and that multi-round extensions let it output videos lasting several minutes while holding characters, scenes and narrative pacing.

The same post says the model accepts up to 30 images, 10 video clips and 10 audio clips as references in one pass, with timestamp-level editing. ByteDance says Seedance 2.5 is rolling out on Jimeng AI and Doubao Pro, with API access coming through BytePlus ModelArk, and that the model still has room to improve on the physical plausibility of complex motion. It ties Kling 4.0 and Wan 3.0 at 30 seconds; the launch page states no price. It places second on control, behind Kling 4.0's larger reference and keyframe set.

1. Kling 4.0 (Kuaishou)

Kuaishou's Kling 4.0 generates up to 30 seconds in a single pass and pairs the longest generation here with the most control over it. Kuaishou's launch release and the model's own user guide both state native single-pass generation of up to 30 seconds, with up to 15 combined reference assets, including up to 10 images, 5 videos and 7 subjects, and up to 10 keyframe inputs.

The guide lists output at 720p, 1080p or 4K, stereo audio, a 21:9 aspect ratio, and input prompts up to 8,000 tokens. Kuaishou says the video-extension feature will soon allow multiple extensions of up to two minutes, which would carry a clip past the single-generation limit, and that a faster Kling 4.0 Flash mode spans 3 to 20 seconds at 720p. Kuaishou's press release says Kling 4.0 launched on October 15, 2026. It takes the top spot because, among the three models tied at 30 seconds, it states the richest set of controls for steering that one generation.