Seedance 2.0 and Gemini Omni Flash

AI video in September 2026 is audio-native, with four names that matter (Seedance, Kling 3.0, Veo 3.1 and Gemini Omni Flash) and one that ended: Sora 2's API shut on 24 September 2026, five months after the Sora app and web closed on 26 April. Veo 4 was not announced. This is a dated letter, written on 14 September, not a model-by-model leaderboard; the update below covers 15 September to 8 October, and the field guide is best AI video models after Sora.

September 2026 was an audio-native video market with four names that still mattered and one clock that actually expired. Seedance 2.0 (ByteDance), Kling 3.0 (Kuaishou), Veo 3.1 (Google), and Gemini Omni Flash sat on one side of that clock. Sora 2's API sat on the other.

Update, 8 October 2026

What changed after the letter below was written, plus one correction:

Sora's API ended on schedule. It shut on 24 September and has returned 410 Gone since the 25th. OpenAI's deprecations table, read again on 8 October, still lists --- as the replacement and has added no video model. Versely removed its Sora 2 rows. See OpenAI named no Sora successor and Sora 2 alternatives.
Kling 4.0 entered early access on 30 September. Kling's announcement describes native 30-second generation, up to 10 keyframe images, up to 15 reference assets and two-channel stereo, with a wider rollout "coming in October"; Kling 3.0 tops out at 15 seconds. Kling 4.0 Flash opened to Ultra Yearly subscribers on 28 September. Kling 4.0 is not a Versely row yet, so the Kling 3.0 and O3 rows below are still the ones to run. Sources: Kling's announcement and its 4.0 vs 3.0 comparison.
Correction on Omni. Google's current Omni model is 1.1 Flash, released on 27 August, before this letter was written. It extends a clip by up to 10 seconds at a time to a 40-second total. The Omni paragraph below describes the June model that Versely still carries; see Omni 1.1 Flash: what changed. Google's Gemini API docs still list Veo 3.1 as the Veo model.
The cheap and long rows are in the catalog. Wan 3.0 starts at 6 credits and Seedance 2.5, with its native 30-second take, at 34.
Stills partners moved. GPT Image 2.5 arrived on 8 September and Google's Nano Banana 2.1 reached general availability on 6 October.

The cross-vendor shift of the quarter is still the same one the trackers have been writing all year: audio-native generation. Seedance 2.0, Veo 3.1, and Kling 3.0 all produce synced audio in one pass; Veo is the one reliably doing 48 kHz dialogue (Magic Hour tracker; wavespeed). If your pipeline is still silent-generate-then-TTS-then-lipsync by default, you are running last year's assembly line on purpose. That assembly line still wins when you need a specific cloned voice or a one-line recut without regenerating the picture. It is no longer the default.

The four that matter, plus the one that is ending

Seedance 2.0. Unified multimodal input (text, image, video, audio; up to 12 references), phoneme-level lip-sync in 8+ languages, up to 15 seconds per generation, native 480p/720p in the technical report with 4K offered on some hosts (fal; Atlas Cloud; Morphic). On Versely the row is Seedance 2.0 (and a fast reference-to-video sibling). Quote credits on that row, not host USD. Seedance 2.5 exists as a longer single-shot line (up to 30 seconds on the catalog row). 2.0 was still the name in the "four that matter" list because of the reference stack and the language lip-sync, not because 2.5 failed to ship.

Kling 3.0. Released February 2026, native 4K at 3840×2160, better physics and camera motion in the February write-ups (pixverse.blog; Magic Hour). In the Versely catalog the 4K text-to-video row is Kling Video V3 4K; Kling O3 Pro is the O3 Pro line. This letter will not pretend those slugs are the same row. The market name in press is Kling 3.0. Pick the catalog row that matches the job (4K, turbo, O3 Pro, motion control), not the press name.

Veo 3.1. Current Veo. Native audio, official 4K tier, short clips (4 / 6 / 8 seconds on the Versely row). Always-on dialogue is the job it keeps winning. Veo 3.1 is that row. Veo 4 was not announced as of mid-2026, and still is not. Write "what we know." Do not write "released" (wavespeed on Veo 4; evolink).

Gemini Omni / Omni Flash. Gemini Omni (Google I/O, 19 May 2026) is any-to-any: text, image, audio, video in; editable video out (TechCrunch; Google Omni docs). Gemini Omni Flash launched 30 June 2026 as conversational video editing; the recommended stills partner in that write-up is Nano Banana 2 Lite (The Next Web). On Versely, Gemini Omni Flash Edit is the video-to-video edit row; Gemini Omni Video is the generate row. Third-party USD-per-second figures exist in those articles; this page will not recast them as Versely prices. Credits on the row.

Happy Horse, Wan 2.7, Runway Gen-4.5, LTX, Hailuo, Pixverse, Vidu, Flux video rows — they are in the catalog and they are in the field guide. They are not this letter's four. A dated state-of is allowed to be unfair about attention.

Sora's API ended on 24 September

App and web had already died on 26 April 2026. The API (sora-2, sora-2-pro, all snapshots) shut on 24 September 2026 (OpenAI help centre; OpenAI community thread). That was the only hard clock in this letter. If you still had code on that endpoint, the migration plan is the checklist. Since the 24th, "Sora" is a history word in a production meeting.

Files you already downloaded remain files. Prompts that only ran against Sora are not a library; they are a bet on an endpoint. Map jobs, not looks: references and languages to Seedance, always-on dialogue to Veo, 4K and motion to the Kling rows, "edit this file" to Omni Flash. There is no public grace window in the shutdown notice. There is no announced Sora 3 to wait for.

Do not let the shutdown meeting eat the rest of September. The four names above are the market. Sora is the calendar event.

Audio-native, any-to-any, and the pipeline that actually changed

Two pipeline changes are real in September, and they are easy to mash into one slogan.

Audio-native means the video model emits picture and sound together. Dialogue, room, lips. Veo's 48 kHz dialogue is the quality marker the trackers keep repeating. Seedance's phoneme-level lip-sync in 8+ languages is the localisation marker. Kling 3.0 in the same breath. The native audio explainer is the mechanics page. This letter's point is adoption: if your default is still a silent generate, you are choosing extra passes.

Any-to-any / Omni means the edit is becoming a conversation on a file you already have, not only a text-to-video roll of the dice. Omni Flash is the named June 30 line for that. Image-to-video is not dead. It is no longer the only respectable pipeline. A still from Nano Banana 2 into an Omni or Veo or Seedance job is a September-normal path. A week of only text-to-video because "that's what we set up in 2025" is a habit.

What did not change: physics still fails, text-on-screen in a generate still fails, product labels still warp, hands still lie. Direction is still the edge because polish is cheap (OpusClip; Hedra; Pictory). 11× more video per month is a workflow number, not a quality number. Shorts at 70 billion daily views will absorb files that hold. It will not absorb a farm because the audio is native.

Platforms lengthened caps (Shorts 3 minutes, Reels 20, TikTok 10). Native-audio models mostly do not fill those caps in one generate. Seedance 2.0 tops out at 15 seconds in the sources above; Veo 3.1's Versely row is 4–8 seconds; Omni Video lists 4–10. Longer objects are still an edit: extend rows, cuts, interviews you actually recorded. Do not wait for a 20-minute native-audio generate because Reels allows 20 minutes.

What a production meeting should decide this month

If you were on the Sora API, you are already off it. The 24th was a hard date. Map jobs to Seedance (references, languages, 15s), Veo (dialogue), Kling (4K / motion), Omni Flash (edit the file you have). Do not map "Sora look" as a job. Looks are not portable. Jobs are.

If you are still storyboarding silent, stop as the default. Write sound in the prompt. Keep TTS+lipsync for the cloned-voice exception.

If you need a stills partner for Omni Flash, use a fast image row, then edit. Nano Banana 2 is the Versely still row in that family. Do not invent a Lite slug if you are linking a catalog page you have not grepped; the named page here is Nano Banana 2.

If the brief is UGC, do not ask a cinematic 4K row to pretend to be a bathroom. Overlay a talking head on a real plate. Native audio on a fake kitchen is how you buy polish you did not want.

If the brief is EU or ads, disclosure is part of the edit. Article 50 has applied since 2 August 2026. Native audio does not exempt a realistic generated person.

If someone asks about Veo 4, say it is not announced. If someone asks about Sora 3, say it is not announced. September is not a rumour meeting.

Credits, not USD, on Versely invoices. Host blog posts will keep printing dollars per second. Those are their numbers.

The catalog will move again before the next dated letter. This one is pinned to 14 September 2026, ten days before the Sora API went dark, with audio-native as the default and Omni Flash as the edit path that did not exist in the spring field guide. The update section above carries it to 8 October.

Run the row the job needs in the AI video generator. Sign up if the missing piece is that row sitting next to captions and the next model when this letter is stale.

FAQ

Is Kling 3.0 the same as Kling O3 Pro in the catalog?

No. Press says Kling 3.0 for the February 2026 4K-native line. The catalog splits V3 4K, turbo, O3 Pro, motion control. Pick the row. Do not brief "Kling 3.0" to a tool that only knows slugs.

Will Omni Flash replace Veo 3.1?

They are different jobs in September. Veo 3.1 is a generate with always-on dialogue. Omni Flash Edit is conversational video-to-video. You can hold both. Replacing one with the other is a brief error unless your job was always "edit this file."

What happened on 24 September?

Sora 2 API calls stopped working, and they have returned 410 Gone since the 25th. Files you already downloaded remain files. Prompts that only lived against that endpoint need a new row. There was no grace window in the public shutdown notice this letter used.

Is there a Kling 4.0 yet?

It is in early access. Kling announced Kling 4.0 on 30 September 2026 with a wider rollout due in October. It is not a Versely row yet; Kling 3.0 and O3 are.