September 2026 was the month the AI video market lost its most famous name and gained its longest-clip contender. Here is what happened, and what it changes if you build products on video APIs.

1. OpenAI switched off the Sora 2 API (September 24)

Six months after announcing the deprecation, OpenAI shut down the Videos API and every Sora model alias, including sora-2 and sora-2-pro. The consumer Sora app had already closed on April 26. OpenAI's documentation lists no replacement model, so Sora integrations have to move to another lab.

What it means: the biggest single-vendor risk in AI video materialized. Veo 3 Fast, MiniMax H3 and Kling 3 are the usual replacements. Full guide: How to migrate off the Sora 2 API.

2. Kling 4.0 Flash opened to early access (September 28)

Kuaishou announced Kling 4.0 and gave a restricted group of subscribers access to its lighter Flash tier: 3 to 20-second clips at 720p. The full Kling 4.0, announced for October, promises up to 30 seconds, up to 4K, 10-bit HDR, native stereo audio and up to 10 keyframes per generation. There is no general developer API yet.

What it means: 30-second single generations are becoming standard at the top end, after Seedance 2.5 and Wan 3.0. Details: Kling 4.0 Flash, what developers should know.

3. Gemini Omni Flash preview endpoint retired (September 30)

Google's Gemini Omni Flash, unveiled at I/O in May and generally available since August 27 as gemini-omni-1.1-flash, closed its preview endpoint at the end of the month. Anyone still calling the preview ID had to move to the stable model. Video output is token-billed, roughly $0.10 per second at 720p. Still no Veo 4: Veo 3.1 remains Google's dedicated video line. Read Gemini Omni Flash explained.

4. The summer releases settle in

Several models launched in July and August became production-ready during September:

  • Seedance 2.5 (ByteDance): released July 31, public API since August 7. 30-second clips with native audio and up to 50 reference inputs. Our overview.
  • MiniMax H3 / Hailuo 3: launched July 31, base weights published August 3 under a license that excludes local deployment in the US, EU, UK and South Korea. The license catch.
  • Wan 3.0 (Alibaba): public beta August 6, release August 24, 30-second clips and document inputs, API-only with no open weights. What it is and is not.
  • Grok Imagine Video 1.5 (xAI): generally available since June, with a July 31 update adding native 1080p, text-to-video and voice references; priced at $4.20 per minute of video.

5. Prices keep falling at the low end

Entry tiers now start around $0.03 to $0.05 per second (Veo 3.1 Lite at 720p, Seedance 2.0 Mini, Wan at 480p), while premium tiers with audio still reach $0.40 per second or more. The spread between cheapest and most expensive option for the same 10-second clip is now more than tenfold. The full table is in our October 2026 pricing comparison.

  • Kling 4.0 full launch and whether a developer API follows quickly.
  • Longer clips everywhere: 30 seconds is becoming the new ceiling for flagship models.
  • Native audio by default: Veo, MiniMax H3, Seedance 2.5, Wan 3.0 and Grok Imagine all generate sound; silent models are starting to look dated.
  • Open weights with strings attached: MiniMax's regional license and Wan 3.0 staying closed suggest the open era of video models is narrowing.

We update the model catalog as new models become available through stable APIs.

Frequently asked questions

What happened to Sora in September 2026?
OpenAI shut down the Sora 2 API on September 24, 2026, after a six-month deprecation period, with no replacement model.
When does Kling 4.0 launch?
Kling 4.0 Flash entered limited early access on September 28, 2026. Kuaishou has announced the full Kling 4.0 for October 2026.
Has Google released Veo 4?
No. As of early October 2026 Google has not announced Veo 4. Its newest video model is Gemini Omni Flash, alongside the Veo 3.1 family.
Which AI video models generate 30-second clips?
Seedance 2.5 and Wan 3.0 already do, and the full Kling 4.0 has been announced with up to 30 seconds.

Sources