Multimodal and Video Generation Brief, Oct 9

Video got longer, images cheaper, rules tighter

Five conclusions

1

Clip length is the new battleground in video

Kling 4.0 went GA on Oct 15 with 3–30 s single takes, 15 references and 10 keyframes, matching Seedance 2.5. VideoGen claims ten-minute films, but only private partners can use Model 1 and nothing is independently tested.

2

Image models compete on editing, layout and price

Black Forest Labs opened FLUX 3 Image Klein (5.6B, bounding boxes, native 4K). Nano Banana 2.1 cuts a 2K image from $0.12 to $0.0593 in fal's test. The top three arena models sit within 20 Elo.

3

Voice agents collapse into single models

Eleven Realtime (Oct 20) replaces the speech-recognition, language-model and text-to-speech chain; v1 Max answered 98.8% of Big Bench Audio questions in Artificial Analysis's test.

4

Music is licensed and policed at the same time

Spotify licensed AI remixes with Believe and pulled 50M+ AI-spam tracks from recommendations on Oct 20; YouTube Music labels mainly AI-made songs since Oct 14.

5

Rights exposure is measurable; EU marking is close

All 15 models in Vn.ai's stress test drew protected characters, Midjourney v8.1 in 37% of prompts. Generative systems on the EU market before Aug 2 must add machine-readable marks by Dec 2.

The lesson

Buy video by the second, images by the licence, and ship nothing unmarked after December 2.

$5.10

A 30-second 1080p Kling 4.0 take at reseller rates (third-party price)

$5M

Annual revenue cap for free commercial use of FLUX 3 Image Klein

What changed: the change log

Date

Model / product

Change

Status

Sep 24

OpenAI Sora

API shut down; the app closed on Apr 26

Retired

Sep 30

HeyGen Video

MiniMax H3 post-train at $0.01/s through October

GA

Oct 6

Nano Banana 2.1

GA on Gemini 3.6 Flash; Nano Banana 2 deprecated

GA

Oct 12

AA-3D

Benchmark for interactive 3D scenes built by models

Released

Oct 14

YouTube Music

Labels on songs made mainly with generative AI

Rolling out

Oct 14

Vn.ai report

All 15 tested models drew protected characters

Published

Oct 15

Kling 4.0

3–30 s single takes, Omni Reference; official API coming soon

GA

Oct 15

FLUX 3 Image Klein

Open weights; free commercial use under $5M revenue

Released

Oct 16

MOSS-Image

Technical report; 5B layout model, Apache-2.0 weights

Released

Oct 20

Eleven Realtime

Speech-to-speech v1 Mini and v1 Max

GA

Oct 20

VideoGen Model 1

Multi-shot videos up to 10 minutes; private partners

Announced

Oct 20

Spotify × Believe

AI remix tool licence; 50M+ AI tracks out of recommendations

AnnouncedIn effect

Oct 20

AFM

Human Made, Human Paid; label talks resume Nov 3

Announced

Dec 2

EU AI Act Art. 50

Machine-readable marking for systems on the market before Aug 2

In effect Dec 2

Kling 4.0 makes 30-second takes the ceiling; long form is still a promise

Launched Oct 15 after a Flash early access from Sep 28. Kuaishou frames the release around steadier large-scale motion, native single takes of up to 30 seconds and Omni Reference.

3–30 s per take, twice Kling 3.0's cap, with dialogue, effects and music in two-channel stereo.
Omni Reference: up to 15 assets (10 images, 5 videos, 7 elements) and 10 keyframes.
Announced, not shipped: 4K, 21:9 and 10-bit HDR; Kling's API packages do not support 4.0 yet.
Plans run $8.80–$544.99 a month; credits per 4.0 video are not published.

Clip cost at reseller rates (UnificAlly)

$0.1134/s at 720p and $0.1701/s at 1080p; third-party prices, not Kuaishou list prices.

Image models compete on editing, layout and price

Cost of one image (fal, Oct 7)

Nano Banana 2.1 halves the 2K price but saves only 16.6% at 4K. Sources disagree on 4K token counts (2,520 vs 3,780).

FLUX 3 Image Klein

ReleasedOct 15, open weights

Size5.6B parameters

Runs on8 GB VRAM (INT8)

Speed~4 s per 1MP, RTX 5090

APIfrom $0.009 per 1K image

Free commercial use only below $5M annual revenue; larger firms need a licence or the API. Quality claims come from BFL's own tests.

Nano Banana 2.1 (GA Oct 6)

Built on Gemini 3.6 Flash with better text and infographics, up to 14 references, search grounding and thinking levels. Billing moved to $30 per million output image tokens. Google's own preference test: 1050 vs 990 (Nano Banana 2) vs 935 (Pro); small text and left-right placement remain weak.

The layout race

GPT Image 2.5 (Sep 8, Flare and Sunburst variants) remains the reference with 16 references and 4K output. MOSS-Image, an Apache-2.0 model described on Oct 16, takes JSON layouts and reaches 96.3% text coverage on LongLongTextBench, the best among open-weight models.

Voice agents become one model; 3D gets a benchmark

Eleven Realtime (Oct 20)

Cascaded voice agent versus one speech-to-speech model

Two speech-to-speech models replace the usual chain of recognition, turn-taking, language model and voice. 90+ languages and 17,000+ voices; the API is OpenAI Realtime-compatible, so switching means a new endpoint and model name.

Independent scores (Artificial Analysis)

AA-3D: scenes that language models build

Launched Oct 12. Models work as agents in an offline sandbox with Blender, physics and CC0 assets, turning a written brief into an interactive three.js page within three hours or 250 turns; a panel then votes blind. 55 tasks, from a rotating black hole to a playable pinball table.

Top 4 tied

Confidence intervals overlap

$11.62–$14.93

Cost of one scene, top models

Music is licensed and policed at the same time

Suno valuation

$400M Series D led by Bond Capital in June; about $300M annualised revenue and 2M paid subscribers by early 2026.

Licensed lane and spam lane

50M+

AI-spam tracks out of Spotify recommendations

75M+

Tracks removed from Spotify altogether

80,000+

AFM members behind Human Made, Human Paid

Spotify × Believe (Oct 20): licence for the AI covers and remixes tool, after UMG, Merlin and Kobalt; Sony and Warner have not signed. A paid Premium add-on that pays original rights holders; no launch date.
Labels: YouTube Music marks songs made mainly with generative AI; Spotify's "AI Persona" profiles are excluded from personalised recommendations.
AFM (Oct 20): consent, credit and compensation before label talks resume on Nov 3.

Rights exposure is measurable and EU marking is six weeks away

EU AI Act Article 50 timeline

Aug 2, 2026

Article 50 applies

Providers mark outputs machine-readably; deployers disclose deepfakes.

Dec 2, 2026

Older systems must mark

Grace period ends for generative systems on the market before Aug 2.

Dec 2, 2027

High-risk obligations

Moved by the Digital Omnibus, in force since Jul 27, 2026.

The voluntary Code of Practice counts as an adequate way to comply and describes three layers: signed metadata, an invisible watermark and fingerprinting.

Vn.ai: every tested model drew protected characters

Vn.ai's first quarterly stress test (September research, reported Oct 14) covered 15 popular models. Every one generated at least one copyright-protected character, even when the prompt did not name it. Midjourney v8.1 was worst, with brand-unsafe characters in 37% of prompts; Seedance 2.5 improved after cease-and-desist letters, while MiniMax H3 still produced violating videos. Disney and NBCUniversal's suit against Midjourney is ongoing and Warner Bros. Discovery is also suing.

Thomson Reuters v. ROSS: training data has a market

On Sep 29 the Third Circuit held that Westlaw headnotes are copyrightable and that using 2,243 of them to train a competing, non-generative legal-research tool was not fair use. The court called it "no more than an ordinary copyright case" and recognised a developing market for licensing training data; generative training questions remain open.

Platforms label AI music now

YouTube added AI music that is the main focus of a video to its disclosure rule (first seen Oct 3), and YouTube Music began labelling songs made primarily with generative AI on Oct 14, using partner disclosures and its own detection; the main YouTube app follows later.

Choose by clip length, licence and marking readiness

Action checklist

0 / 5 done

Price video per second and per clip; treat Kling 4.0's 4K and HDR as announced until the API lists them.

Move 1K–2K batch image work to Nano Banana 2.1 or Klein; re-measure 4K before switching.

Read every open-weight licence: Klein is free only under $5M revenue; MOSS-Image is Apache-2.0.

Add character and brand checks to review, and keep prompts and outputs for audit.

Confirm every generator you ship into the EU adds machine-readable marks before Dec 2.

Dates to watch

Oct 31
HeyGen Video's $0.01/s launch price ends
Nov 3
AFM resumes talks with Universal, Sony, Warner and Hollywood Records
Dec 2
EU marking for generative systems on the market before Aug 2
End of 2026
VideoGen distributes its $1M credit pool

Divrite Research · Hot Topics · Multimodal and video generation, issue for Oct 9–22, 2026. Vendor-reported figures and third-party prices are labelled as such.