Hot topic · Multimodal and video generationIssue: Oct 23–Nov 2, 2026

Video clips converged on 30 seconds as access, price and rules moved

The one genuinely new item was Genie 4, an interactive world model; meanwhile model IDs, price lists and labelling rules changed more than clip length or rank did.

How to read this page: figures come from public vendor releases, benchmark boards, court reports and press coverage; each keeps its status (vendor claim, reported, estimate, allegation, upcoming) and date. It is a dated fact brief, not a ranking, legal advice or investment advice.

30 s
Single take, Kling 4.0 API (vendor spec; on fal Oct 23). Kling 3.0: 15 s
18 Elo
Gap across the top five text-to-video models (Artificial Analysis, undated)
14.2x
Price spread, $2.40 to $34.12 per minute, four models with an API (undated)
Divrite Research · Status as of Nov 3, 2026
Hero visual for the multimodal and video generation brief
274 songs
Presley catalogue owners v. Anthropic, filed Oct 23; $48.635M statutory maxima pleaded
Allegation; pending
40M+
Higgsfield registered users claimed on Oct 23 (30M at the August round)
Company statement; unverified
Oct 22
Gemini API Veo 3.1 preview IDs shut down, one day before the window
Reported; in effect
Nov 17
Google Cloud veo-3.1-generate-001 retires on this date or later; a floor, not a commitment
Upcoming
01

Capability news was thin while access, price and rules moved

Five conclusions from Oct 23 to Nov 2, 2026, then the dated change log. Rows dated before Oct 23 are context.

Five conclusions

1

Few launches: clips converged on 30 s with audio, and the new item is an interactive world model

Only two video events fall in the window: Genie 4 (research preview, Oct 27, trusted testers, waitlist) and Kling 4.0's API on fal (Oct 23, single takes up to 30 s against Kling 3.0's 15 s). Seedance 2.5 and Wan 3.0 also claim 30 s but launched earlier (context). Hi3D V1.0 open weights (Oct 23) is the only dated downloadable release; Midjourney V9 is not out.

2

For builders, access moved more than capability

Google's Gemini API Veo 3.1 preview IDs were reported shut down on Oct 22, one day before the window, with Gemini Omni Flash, itself a preview, as replacement. Google Cloud's veo-3.1-generate-001 retires on Nov 17, 2026 or later, a floor, not a commitment. Kling 4.0 is reachable through hosts such as fal; no Kling-owned API page is in the material.

3

Price, not rank, separates the leading video models

On Artificial Analysis (undated snapshot) the top five text-to-video models sit within 18 Elo (1174 to 1156), yet the four with a public API list at $2.40 to $34.12 per minute, a 14.2x spread. On Runway's rate card a 10-second 1080p clip costs $2.00 on Wan 3.0 and $6.80 on Seedance 2.5; Seedance 2.5 at 1080p is $0.68 per second at Runway and $0.90 at WaveSpeed.

4

Rights holders sue and license in parallel, and no court has ruled on training

Presley catalogue owners filed against Anthropic on Oct 23 over 274 songs, with statutory damages named at $48.635 million: an allegation, not an award. In the Suno cases a magistrate judge allowed a deposition on Oct 26, and Apple's subpoena fight has a hearing on Dec 1. Platform rules stay at labels and provenance, such as the EU AI Act marking duty in effect since Aug 2.

5

Money lands on the distribution layer, and every growth figure is self-reported

Higgsfield added Nvidia and TCV to its Series B extension on Oct 23 with amount and valuation undisclosed, while claiming 40M+ users and $1B+ annualized revenue (company statements, not independently verified). Utopai Studios is reported at $60M on a $1B valuation. Runway both resells models and supplies plugins to Adobe's editors.

The window in one line

The window moved less on capability than on access, price and rules.

Model IDs, price lists and labels are dated facts to track; distribution and licensing, not clip length, are where the contest is shifting.

  • Video events inside the window: 2 (Genie 4, Kling 4.0 API)
  • Dated downloadable releases: 1 (Hi3D V1.0, Oct 23)
  • Licence deals dated inside the window: 0 in the material
  • Court rulings on training: none
Dated inside the window
Oct 23Kling 4.0 API on fal; Hi3D V1.0 weights; Higgsfield round; Presley suit filed
Oct 26ElevenLabs v4 clone_strength; Suno deposition order
Oct 27Genie 4 announced as a research preview

What changed at a glance

Filter

Status shown as reported in the sources; Oct 22 sits one day before the window and is kept as context.

02

Only two video events fell in the window and the new one is a world model

Specs are vendor claims unless stated. Orange-edged cards are dated inside Oct 23 to Nov 2; the rest are context from earlier weeks.

Show

Kling 4.0 timing conflicts

Kuaishou's own page is not in the material, so every date is second-hand.

  • Oct 16: a reseller comparison called the model “fully released” on Oct 15 but listed Kling's own developer API as “coming soon”.
  • Oct 22 (vendor claim): a paid press release says the API “officially launched on October 22”, opening gradually to enterprises and developers.
  • Oct 23, 3:00 PM: fal posted that Kling 4.0 was “publicly available” with API access “open to all customers”. This is the earliest dated public statement and the one followed here.
  • Oct 28: CellCog says the API opened. None of the four is Kuaishou's own.

World model versus clip generator

  • Interaction: prompt, then wait for a clip.
  • Coupling: standalone video model.
  • Stated use: ads, storyboards, reference-driven scenes.

Genie 4 is the only world model in the material, so no head-to-head is possible. With waitlist-only access, builders cannot yet depend on it.

Not found as in-window releases: Runway's changelog (as crawled) tops out at Oct 2 with Seedance 2.5 Draft Mode in Runway Agent. Luma Ray 3.2, Grok Imagine 1.5 and Midjourney's image-to-video are listed with no release date stated. Midjourney's V2 video model is not out: a creator recap posted Nov 1 says video becomes a bigger focus after V9.
03

What builders can still call: retirements, hosts and open licences

No retirement dated inside the window appears in the material; the window's access news is Kling 4.0's API and Hi3D's open-source release.

Deprecation and retirement calendar

Filter

The replacement is a preview

129 days

is how long the Veo 3.1 previews lasted as Google's named Veo 3.0 replacement (Jun 15 to Oct 22).

  • Gemini Omni Flash: public preview since Jun 30, 3 to 10 s at 720p.
  • Veo 3.1 makes 4, 6 or 8 second clips.
  • Google says listed shutdown dates are the earliest possible.
  • Sora (context): per Dead Canary, quoting OpenAI's deprecations page, developers were notified Mar 24, 2026 and the Sep 24 removal named no replacement; OpenAI's own page is not in our material.

Google's own page is missing from our evidence

The Oct 22 date rests on a tracker, a review site and a third-party review that cites Google's documentation, so it is treated as reported.

Google Cloud's Veo 3.1 GA dates from Nov 17, 2025; “Nov 17, 2026 or later” is a floor, not a commitment, and models may stay reachable through the Gemini API. The page shows pay-as-you-go unsupported, provisioned throughput and fixed quota supported, us-central1 only.

Kling 4.0: a paid press release (Oct 23) says the API launched Oct 22; fal's dated post (Oct 23, 15:00) is the earliest verifiable event. ReviewRamble says first-party pricing is unpublished, so treat access as “API available through hosts”.
Where Veo 3.1 survives is open: HeyGen (Oct 29) says Omni Flash replaced it inside the Gemini app; AI Foresight 360 says Google's Veo page still lists it in the Gemini app and Flow. Neither is a Google page.

What to do before the next date

Tap an item to tick it off. Dated items are labelled in effect or upcoming as of Nov 3, 2026.

04

Price, not rank, separates the leading video models

The top five text-to-video models sit within 18 Elo (1174 to 1156) on Artificial Analysis, but the four with a public API list at $2.40 to $34.12 per minute, a 14.2x spread (34.12 / 2.40).

Quality and price snapshot

Reading the boards

  • Ranks overlap. The board's own rank ranges (Wan 3.0 1–3; HeyGen and Seedance 2.5 2–7) overlap, so places 1 to 5 are not cleanly separated.
  • Leadership flips by task. Seedance 2.5 leads image-to-video (1157) while Wan 3.0 is ninth (1111); Wan 3.0 leads video editing (1196 against 1152); HeyGen Video 1 leads the silent text-to-video pool (1143); MiniMax H3 (768p) leads open-weights video (1158). Kling 4.0 is absent from the retrieved boards, which show only Kling 3.0.
  • Image and voice. GPT Image 2.5 Sunburst also leads editing (1199); Nano Banana 2.1 trails it by 46 Elo (1222 - 1176) at 6.3x lower price ($210.7 / $33.6); FLUX 3 Image Klein leads open-weights images (1043, $9.0 per 1k). Speech: Eleven v4 and v4 Turbo tie at 1333 ($80 and $40 per 1M characters); Voxtral 3 TTS scores 1221 at $13.0.

Price spread

14.2x

between HeyGen Video 1 at $2.40 and Seedance 2.5 at $34.12 per minute (Artificial Analysis list price, undated snapshot).

Elo comes from blind pairwise votes (Bradley-Terry); video boards anchor 1000 at Kling 3.0 1080p (Pro).

Price tables

Runway's developer rate card is the only page in the material that prices many makers' video models in one unit (1 credit = $0.01). It is Runway's resale price, undated in the material.

A 10-second 1080p clip therefore costs $2.00 on Wan 3.0 (10 × $0.20), $2.35 on Kling 4.0 at list (10 × $0.235), $4.00 on Veo 3.1 with audio (10 × $0.40) and $6.80 on Seedance 2.5 (10 × $0.68).
05

Rights holders sue and license in parallel, and no court has ruled on training

A complaint is an allegation. The material holds no docket pages, so statuses rest on press reports and law-firm trackers, as of Nov 3, 2026. Nothing here is legal advice.

Litigation and settlements

Filter

Presley v. Anthropic: damages as pleaded

Statutory maxima named in the complaint, filed Oct 23, 2026: 274 works × $150,000 (willful infringement), × $25,000 (credit removal), × $2,500 (circumvention); total $48.635 million. An allegation, not an award. Anthropic: “We disagree with the claims and intend to defend this case fully.”

What the complaints allege and what is at stake

  • Presley complaint: licensed lyrics were scraped from Common Crawl, Anthropic removed credit information in data cleaning, and Claude's lyric filters are “readily circumvented by rephrasing the prompt” (allegations). Anthropic had not answered as of Oct 25.
  • Reservoir Media: 1,050 × $150,000 = $157.5 million on the main claim; adding $25,000 and $2,500 per work gives $186.375 million (allegation ceiling).
  • UMG and Sony: 60,202 × $150,000 ≈ $9.03 billion, a theoretical ceiling set against a $5.4 billion Suno valuation in June, according to a video commentary.
  • Procedure: on Oct 20 Suno told the Boston court it “has not deleted all” training recordings; on Oct 26 (window) a magistrate judge let the labels depose a former Menlo Ventures principal. GEMA won in Munich on Jul 31 (context); Suno can appeal.
  • Gap: no filing or docket event for the Disney-led studios' suit against MiniMax (Hailuo) is in the material.

Licensing deals and training-data policy

No licence deal dated inside the window appears in the material; the map below is context, and deals are announced terms, not rulings.

Udio (context): Universal settled in October 2025 and Warner in November 2025. A “walled garden” where nothing leaves the platform; downloads off since October 2025, with Warner, Merlin and Kobalt deals signed (TechieHub).
Suno (context): Warner settled on Nov 25, 2025; Suno committed to train only on licensed works from 2026. Suno v6 (Sep 9) was built with Warner, BMG and Believe; UMG and Sony call it “the fruit of the same poisoned tree”. Universal talks stalled in April 2026.
ElevenLabs Music (according to): TechieHub and BuildMVPFast describe it as licensed-data, with Merlin and Kobalt deals. No ElevenLabs release is in the material.
Others (context): Klay signed all three majors on Nov 20, 2025. LANDR's Fair Trade AI pays artists 25% of net licensing revenue plus a $1 million advance fund, expanded July 2026.

Suno's terms give it a perpetual, royalty-free licence to train on user creations with no opt-out. The window's contribution is fair-use evidence: Suno wants Apple's AI-clause contracts for the market-harm factor. The only US appellate ruling in the material, the Third Circuit's Sep 29 Thomson Reuters v. Ross (context), rejected fair use for a non-generative tool and reserved the generative question.

Platform labelling and disclosure rules

Provenance. Google says it embeds SynthID and C2PA markup in all images and videos from its AI tools. The SynthID Detector portal opened on Oct 7 (context); Gemini promoted its “Is this AI generated?” check on Oct 27 (window), limited to one file up to 100 MB. C2PA's index pointed to version 2.4 on Oct 2. NIST reports that most watermarks are removable.
Still unresolved. Fair use for generative training has no US appellate ruling. Anthropic had not answered the Presley complaint as of Oct 25. Upcoming: Midjourney dispositive motions by Nov 23, Apple hearing Dec 1, EU marking deadline Dec 2. A creator reported on Oct 25 an uneditable AI label on a podcast with only noise reduction.
06

Money lands on the distribution layer and every growth figure is self-reported

The only new financing in the window is Higgsfield's Series B extension, announced on Oct 23, 2026, with amount and valuation undisclosed. Everything else below is a report or context from before Oct 23.

Funding and valuations

Kling figures differ by source: Kuaishou's filing, as quoted by gradually.ai, gives a $3B subscription limit; press reports give $2.8B or “nearly $3B”. The filing's limit is followed as a ceiling; the $15B is a pre-transaction valuation, not post-money.

Reported valuations, mixed dates and bases

Each bar carries its own basis and date; they are not comparable like for like. Higgsfield's Oct 23 extension valuation is undisclosed and not shown.

What the money signals

Capital is landing on the packaging layer (Higgsfield) and on a studio that borrows open models (Utopai), while every growth number is self-reported.

  • Mixed-date check: Higgsfield's August valuation over its September run-rate is $5.4B ÷ $1B = 5.4x; Runway's February valuation over its September ARR is $5.3B ÷ $200M = 26.5x (Bloomberg, Sep 8, via gradually.ai). Rough ratios, not valuations.
  • Nvidia appears on both Higgsfield and Runway; TMC Insight reads that as spreading bets, an interpretation, not a company statement.
  • Market-size estimates disagree: $946.4M (Grand View Research) and $3.67B (Meticulous Research) against $3.0B for 2026 (LinkedIn market report, Oct 27, 2026). Definitions differ, so none is used as a base.

Higgsfield company statements, not independently verified

40M+
Registered users, Oct 23 (30M at the August round)
$1B+
Annualized revenue, crossed in September ($700M in August)
10x
Revenue under contract, business customers, since June
238
Countries and territories, Oct 23
150M
Generations per month

Run-rate growth is (1,000 − 700) ÷ 700 ≈ 43% between the August round and September: a run-rate, not recognized revenue (estimate). TMC Insight, via Bloomberg, says revenue is “approaching” $1B while the company says it crossed $1B; the company's wording is followed but labelled a vendor claim. Briefs says Higgsfield “added 10 million paying customers”, whereas Higgsfield's own post says users grew from 30M to 40M+; the 10M is read as registered users, not payers. Its enterprise claim also appears as “contracts up 71%” since August; the baselines differ and cannot be reconciled.

Distribution and aggregator platforms

No new reseller deal is dated inside the window; the map is mostly context, but it shows who sells whose model.

Runway buys and sells. Its Aleph model sits inside Firefly's Prompt to Edit, and Runway itself runs inside Adobe's editors.
Aggregators outlive retirements. Per AI Foresight 360, Google shut Veo 3.1 previews on Oct 22 (reported), yet Higgsfield still offers Veo 3.1.
Nvidia is infrastructure too. Higgsfield scaled on Nvidia GPUs from L40 and A40 through Hopper and Blackwell and is moving to Vera Rubin.

Utopai Studios: a different bet

Utopai makes films rather than selling tools, licenses its technology and built its model on a customized MiniMax H3. Artificial Analysis lists Utopai X second on its text-to-video board (Elo 1168, no API available).

20+
Projects, including two TV series (reported)
3,000
North American theaters planned for three features in 2027 (reported)
80–90%
Claimed cut in scene-planning and CGI costs (vendor claim)
$2B–$3B
Term sheet valuation, reported but not completed

Creator economics: credits, prices and monetization

According to AI Foresight 360, Higgsfield plans cost $9, $29 and $79 a month (monthly billing) for 120, 600 and 1,800 credits; Creatorflow lists different tier names and quotes no price (checked July 2026). No vendor pricing page is in the material, so treat the plan prices as unconfirmed. On Runway, per Creatorflow citing the pricing page (Oct 2026), the $15 per month Standard plan includes Kling 3.0 and Gen-4.5. Runway's enterprise offer of two free months for contracts signed by “10/31” has now passed.
07

The contest is shifting to control points, with dated risks ahead

A market moving from model launches to control points: who hosts a model, what it costs per second and what rules apply to its output. Readings of the evidence, not legal or investment advice.

Trends and how strong the evidence is

The strongest signal is price dispersion: the same model costs very different amounts by seller, and the cheapest is not the top-ranked. The weakest is the world-model story, which rests on one vendor post.

What it means by reader

What changed

Credits differ by platform: Runway Max works out at about $0.096 per Gen-4.5 second on annual billing; Higgsfield plans are unconfirmed.

Implication

Compare cost per second, not plan price. Kling 4.0's 20% promo on Runway's developer rate card ends Nov 30 (upcoming).

Where to look

Module 04 price tables; module 06 creator economics.

Key dates

Dates before Nov 3, 2026 are in effect or carry the matching status word; later ones are upcoming.

Filter