The '8-second limit' is one model's cap (Veo 3.1) mistaken for an industry rule. As of August 2026, native single generations run 8-30 seconds — Seedance 2.5 holds the 30s record, FLUX 3 does 20s — and extension chains reach 120s (Sora API), 148s (Veo, 720p-only), and ~3 minutes (Kling, paid, with drift). Real long-form is assembled shot by shot.
Updated August 2026
"AI videos max out at eight seconds" is the most repeated spec claim in the category, and it has been wrong for a while. Eight seconds is one model's per-generation cap — Veo 3.1's — fossilized into a genre-wide rule by a thousand explainers. As of August 2026, native single-generation ceilings actually run from 8 seconds (Veo) to 20 (FLUX 3) to 30 (Seedance 2.5), and extension chains stretch documented maximums to 120 seconds (Sora's API), 148 seconds (Veo), and about 3 minutes (Kling, and Seedance 2.5 per beta reports). The honest answer to "how long can AI videos be" is therefore three numbers per model, not one — the native cap, the extended cap, and the catch attached to the extension — and this page lists all three for every major model.
Maximum video length by model, August 2026
| Model | Native single generation | Extended maximum | The catch |
|---|---|---|---|
| Veo 3.1 | 4, 6, or 8s (1080p/4K are 8s-only) | 148s via +7s extensions, up to 20 times | Extensions are 720p-only, Veo-source-only, on a 2-day storage timer |
| Sora 2 (API) | 4–20s (steps of 4) | 120s via up to 6 chained extensions | API shuts down Sept 24, 2026 |
| Kling 3.0 | 15s (multi-shot: up to 6 labeled shots in one gen) | ~3 minutes via ~4–5s extension steps | Paid plans only; reviewers report drift past ~2 minutes |
| Seedance 2.0 | 2–12s via API (15s on some consumer surfaces) | Via 2.5, below | Ceiling differs by surface |
| Seedance 2.5 | 30s native — the longest single pass on the market | Extend twice officially; ~3min in beta reports | Public July 31, 2026; API GA ~Aug 7 |
| MiniMax-H3 (Hailuo) | 4–15s (15s holds even at 2K) | No documented extension | Pre-H3 Hailuo capped at 6/10s |
| HappyHorse | 3–15s | No documented extension | Physics artifacts reported in longer generations |
| PixVerse V6/C1 | 1–15s, any integer | ~60s practical via ≤15s extend segments | No documented ceiling — ~3 extensions is the realistic limit |
| Wan 2.6/2.7 | 2–15s | Video continuation (2.7) | Open-weight Wan (≤2.2) is fixed at 5s |
| LTX-2.3 | ~10s consistently documented | Extend operation on the Pro endpoint | Some LTX-2 launch coverage cited 20s; no official maximum published |
| FLUX 3 | 20s native at ~720p | Keyframe conditioning + clip continuation | Early Access only as of Aug 2026 |
| Grok Imagine | 1–15s | Video Extend, no documented total cap | Extend's duration parameter counts only the added seconds |
| P-Video (Pruna) | Up to 15s | No documented extension | Draft-tier model by design |
Two things the table shows that the 8-second myth hides. First, 15 seconds — not 8 — is the modal native cap in 2026. Second, the two axes don't correlate: Veo has the shortest native ceiling on the table and the longest documented extension chain, while Seedance 2.5's record 30-second native pass comes with the stingiest official extension allowance (twice).
Where did the 8-second myth come from?
From reading Veo's spec sheet as the industry's. Veo 3.1 generates 4, 6, or 8 seconds per pass, Google's docs are the most-read in the category, and "8 seconds" propagated as the answer for AI video generally. It was never even Veo's whole answer — the same docs describe a scene-extension mechanism that adds 7 seconds per step, up to 20 times, for a true maximum of 148 seconds. The myth survives because most coverage stops at the generation parameter and never reads the extension section.
How do extension chains actually work — and where do they break?
Extension means the model generates a continuation conditioned on the end of an existing clip, then splices it on. Every implementation rations it differently, and the constraints matter more than the maximums:
| Model | Step size | Documented ceiling | Constraints that bite |
|---|---|---|---|
| Veo 3.1 | +7s per extend | 148s total | 720p only — extending forfeits 1080p/4K; source must be Veo-generated; clips sit on a 2-day storage timer that resets on each extend |
| Sora 2 API | Up to +20s per extend | 6 extensions, 120s total | The whole mechanism sunsets with the API on Sept 24, 2026 |
| Kling | ~4–5s per extend | ~3 minutes total | Paid feature; reviewer-reported quality drift beyond ~2 minutes |
| Seedance 2.5 | One continuation pass at a time | Extend twice, officially | ~3-minute totals appear in beta reports — treat as unconfirmed |
| PixVerse | Up to +15s per extend | None documented | Practical consensus lands around 60s before coherence decays |
| Grok Imagine | You choose the added length | None documented | duration = the extension only: a 10s input extended "5s" returns 15s |
The verdict on this table is unglamorous: every chain trades something — resolution (Veo), money (Kling), coherence (PixVerse, and Kling past two minutes), or existence (Sora). Extension is a splice, not a memory; characters and lighting drift a little at every seam, which is why no model advertises unlimited chaining even when its docs set no cap.
Grok's parameter semantics deserve the callout because they're the opposite of what most people assume: on most models a duration setting describes the output, but on Grok's Video Extend it describes only the addition. Budget accordingly.
What's the real answer for videos longer than a minute?
Assembly, not extension. Extension chains inherit compounding drift and hard ceilings; actual long-form AI video in 2026 is made by generating shots at each model's native length — where its quality peaks — and sequencing them. That's how the documented long-form work gets made: Kling's own multi-shot mode structures up to 6 labeled shots inside a 15-second pass precisely because cuts are cheaper than continuity, and the Dor Brothers' 14-minute Seedance-made film "Apex" (February 2026) was assembled from short clips, not one heroic chain.
Practically, that workflow needs two tools: a generator that can produce shots across models, and an editor that joins them. Inside invideo, the AI video generator handles per-shot generation against the full roster above — the agent sequences shots model-by-model, with character and location memory carrying across cuts — and the AI video extender covers the cases where a genuine continuation of an existing clip is the right move. The 15-second ceilings stop being ceilings when the unit of work is the shot, not the video.
Length limits, answered
How long can an AI-generated video be in one generation? Between 8 and 30 seconds on current major models, as of August 2026. Veo 3.1 sits at the short end (8s max), most of the field — Kling 3.0, MiniMax-H3, HappyHorse, PixVerse, Wan 2.7, Grok, P-Video — caps at 15s, FLUX 3 reaches 20s, and Seedance 2.5 holds the record at 30 seconds in a single pass.
Can AI make a full 1-minute video? Yes, three ways: Sora 2's API reaches 120s via chained extensions (until Sept 24, 2026), Veo 3.1 reaches 148s at 720p, and Kling extends to about 3 minutes on paid plans. The more reliable route at that length is assembling multiple native-length shots in an editor.
Why does my extended Veo video drop to 720p? Because Veo's scene extension is documented as 720p-only. Choosing 1080p or 4K locks you to a single 8-second generation; choosing extension locks you to 720p. There is no path to a long and high-resolution Veo clip in one output.
Which AI video model makes the longest videos? By documented single generation: Seedance 2.5 (30s). By documented extension ceiling: Kling (~3 minutes, paid, with drift reported past 2). By documented chain on a per-second-billed API: Veo 3.1 (148s). The right answer depends on which of the three numbers your job cares about.
Do longer AI videos cost more? Proportionally, on per-second and token billing — a 148-second Veo chain bills every second — and extension itself is gated to paid tiers on Kling. Longer generations also fail more, and only some vendors (MiniMax, HappyHorse) don't bill failures.
Is there a model with no length limit? No. Even models whose extension docs set no numeric cap (PixVerse, Grok) degrade in coherence within about a minute of chaining. Unlimited-length claims are a reseller tell, not a spec.
Version history: first published August 2026. Notable movements tracked: Seedance 2.5's 30-second native pass (public July 31, 2026), FLUX 3's 20-second Early Access generations (announced July 23, 2026), and the pending Sora API sunset (September 24, 2026).