Best AI Tools for Video, Music & 3D That Held Up

Last checked 6 November 2026. This is a desk analysis of documented cases, first-party documentation, and press reporting. I did not run my own benchmarks, and I say so wherever a number depends on a vendor or a single outlet. A one-hour test you can run yourself is near the end.

How to read the evidence tags. [V] verified in a first-party source (the company, a regulator, a survey organizer, a published model page). [R] reported by a credible outlet or practitioner, not independently confirmed. [E] vendor claim, estimate, or a single secondary source. Treat [E] as a lead, not a fact.

In this article: The short answer · Eight cases · Tools are rented · Video · Music · 3D · Prompt helpers · Rights · Run your own test · FAQ · Limits

The short answer

There is no single best tool for video, music, or 3D, and last year’s answers have already expired. What you can do is match the job to a tool with a clear current status.

Job	Reasonable starting point	Why	Watch out for
Polished 10-second clip with sound	Google’s Gemini Omni	Google says it replaces Veo in the Gemini app; 10-second clips, native audio, up to 5 reference photos, scene extension, multi-turn editing [V]	Model names and API routes are still moving
Directed shots you iterate on	Runway Gen-4.5	Image-to-video added March 2026; Lionsgate and AMC Networks reported as users [R]	Credits burn fast
High-volume video on a budget	Kling 3.0	Reviewers keep calling it the value pick [E]	Those are roundups, not audited tests
Song drafts and demos	Suno v6	Newest model, built partly on licensed catalogs [R]	Sony and Universal’s suit against Suno is still open [R]
Music you want fewer licensing questions about	ElevenLabs Music or Stable Audio	Both say they trained on licensed data [E]	Those are the vendors’ own claims; read the plan terms
Licensed remixes and covers	Wait for Spotify’s add-on	Launch reportedly planned for Q1 2027 [R]	Not available yet
Quick game props in 3D	Tripo or Meshy	Built around game workflows and rigging [E]	Output is a draft, not a finished asset
3D you run yourself	Microsoft’s TRELLIS.2	Open model, 4B parameters, full PBR output [V]	Needs a serious GPU; read the license

Everything above is a snapshot of one week. The more durable part of this piece is why it will look different by spring, and what to keep when it does.

Eight cases, and how far to trust them

I started from cases rather than tools, because “AI video is amazing” is a mood and “Netflix shipped one shot” is a fact you can check.

#	Case	What happened	What it actually proves	Evidence
1	Netflix, El Eternauta	A building-collapse shot made with generative AI, in about a tenth of the time, per Netflix	AI can unlock a shot the budget couldn’t	[R] (Netflix’s own claim)
2	Sora shutdown	OpenAI closed the app and API on a few months’ notice; Disney’s $1B deal collapsed	A loved product can still be uneconomic for its maker	[V] shutdown; [R] the reasons
3	Veo 3.1 to Gemini Omni	Preview API models retired 22 Oct; replacement uses a different API	Prompts port between models; code and file formats don’t	[V] notice; [R] migration detail
4	Deezer’s upload data	AI tracks passed half of daily uploads but are 1–3% of streams	Volume isn’t demand, and most AI streams were fraud	[V]
5	Suno and Udio vs. the labels	Lawsuits turned into licensing deals; Udio became download-less	“Legal” now depends on whose catalog trained the model	[R]
6	GDC 2026 survey	52% of game professionals say gen AI hurts the industry	Reception is part of your risk, not just capability	[V]
7	RetroStyle Games	Tripo gives a hi-poly starting mesh; shipped assets are finished by hand	AI 3D is a head start, in a studio that says so	[R] (first-party)
8	Tavern-scene study	AI-assisted scene took 238 min vs 716 min by hand, with heavy cleanup debt	Faster first pass, slower last mile	[E] (one small study)

Read the rightmost column before you quote any of these. Case 1 is a company talking to investors. Case 8 is a single practice-led study in a small journal. They are useful, and they are not proof.

Tools are rented

Most “best AI tools” lists rank products as if they were stable. 2026 didn’t cooperate.

OpenAI announced on 24 March that it was discontinuing Sora in both the consumer app and the API, saying only that timelines would follow (Engadget, VentureBeat) [V]. Disney had signed a three-year deal in December 2025 with a planned $1 billion equity investment and licensing of more than 200 characters. It fell apart, and reporting says no money had changed hands (BusinessDay, Exchange4media) [R].

Google did something quieter. It announced Gemini Omni on 19 May and says Omni will replace Veo in the Gemini app (Google) [V]. In the developer API, the Veo 3.1 preview models were scheduled to shut down on 22 October, with gemini-omni-1.1-flash named as the replacement (Gemini API deprecations) [V], and a developer filing an issue on an open-source project noted the replacement uses a different API call pattern, so it isn’t a simple model-name swap (GitHub issue) [R]. Third-party trackers say Veo 3.1 remains on Google’s enterprise platform with an earliest retirement date of 17 November (innfactory) [E], which I could not confirm on a Google page.

Even outside video, the 3D side churned: Meshy retired its meshy-5 API model on 10 October, and one review reports Luma’s Genie text-to-3D was retired on 1 January (Frame Sixty) [E].

Sources: Engadget, Google’s Gemini pages and deprecation notice, Music Business Worldwide, The Decoder, Gizmodo, Frame Sixty. Veo 3.0 date per the Gemini deprecation schedule as relayed by Layer3 Labs [E].

So the useful question isn’t “which tool is best.” It’s “which parts of my work survive when the tool goes away?” In every case above, the answer was the same: your prompting habits, your reference material, your edit decisions, and your records. The model is rented.

Video

The one shot that worked

The cleanest positive case I found is also the least glamorous. Netflix co-CEO Ted Sarandos said on an investor call that the Argentine series El Eternauta used generative AI for a building-collapse sequence in Buenos Aires, completed about ten times faster than traditional VFX tools would have allowed, and that it would not have been feasible within the show’s budget (Engadget, Korea JoongAng Daily) [R]. The team worked with Eyeline Studios, a production group inside Netflix (Business Recorder) [R], and press accounts place the AI footage in two brief shots in episode 6.

Read that carefully, because it’s easy to over-read. The “10x” is Netflix’s own claim, made to investors, about one sequence. No one has audited it. What’s solid is the pattern: a shot the production couldn’t otherwise afford, a purpose-built pipeline, human review, and a tiny slice of screen time. Nothing in it says AI can carry a whole episode.

Runway’s partnerships follow the same shape. Lionsgate and AMC Networks are reported as Gen-4.5 users for filmmaking and marketing material [R]. Runway also says that in its own study, more than 90% of participants couldn’t reliably tell five-second Gen-4.5 clips from real footage (release notes via Releasebot) [E]. It’s the vendor measuring its own model, so I’d use it only as a hint that detecting AI footage by eye is no longer dependable.

The tools, today

Tool	Status on 3 Nov 2026	Evidence
Gemini Omni (Google)	Replaces Veo in the Gemini app; 10-second clips, native audio, up to 5 reference photos, scene extension, video-to-video and multi-turn editing	[V] Google
Gemini Omni 1.1 Flash (API)	Released 27 Aug; 360p drafts, upscale to 1080p/4K; about 40 seconds by chaining extensions	[R] The Decoder; [E] the 40 s figure
Veo 3.1	Gemini API previews shut 22 Oct; enterprise availability reported to continue for now	[V] notice; [E] the rest
Runway Gen-4.5	Flagship; image-to-video added March 2026	[E]
Kling 3.0 (Kuaishou)	Released 5 Feb 2026; widely recommended for value	[E] roundups
Seedance 2.0 (ByteDance)	Newer entrant in 2026 roundups	[E]
Sora	Discontinued	[V]

I’m deliberately not naming a winner. The roundups disagree with each other, usually because they test different clip types, and several are published by companies that sell access to these models. One hands-on comparison scored Kling’s Pro tier at 91/100 against 63 for Veo 3.1 (Tech Insider); others rank Veo first for realism. Both can be true.

What a clip costs, and why it should change how you prompt

Matthias Bastian at The Decoder reported Google’s per-second prices for Omni 1.1 Flash: $0.03 at 360p, $0.10 at 720p, $0.15 at 1080p, and $0.30 at 4K (The Decoder) [R]. Google’s own pricing page states about $0.10 per second at 720p, billed by tokens (5,792 output tokens per second of 720p) (Gemini API pricing) [V]. The other tiers are reported rather than confirmed by me, so check the live page before you budget.

The ladder is ten to one from draft to final. And the price on the label isn’t the price of a usable clip. If only some generations are keepers, the effective cost is the price divided by your keeper rate. These keeper rates are illustrative, not measured:

Keeper rate	720p clip	4K clip
1 in 2	$2	$6
1 in 4	$4	$12
1 in 8	$8	$24

That’s the case for drafting cheap and finishing once. Google positions 360p drafts as about a third of the 720p cost and up to 60% faster, per the same report [R]. Log your own keeper rate for a week, and you’ll know your real price.

Prompting video: what the primary documentation says

Google’s own prompt guide for its video models is the best primary source I found (Google prompt guide) [V], and its best-practices page adds two details most blog guides skip (best practices) [V]:

Google suggests using a text model to help: show it a reference image and ask it to act as an expert video prompter and write a comprehensive prompt for what you want.
Veo’s built-in prompt rewriter expands your prompt, adding scene description, camera motion, and sound cues, and a rewritten prompt is returned only when your original is under 30 words (rewriter docs) [V]. That documentation covers Veo on Vertex AI, so check whether the same behavior applies to the model you’re using. If it does, a short prompt is partly written for you, and if you want exact control, you either write longer or switch the rewriter off.

Practitioner guides built on Google’s guidance converge on the same checklist: subject, context, action, style, camera motion, composition, ambiance and, because current models generate sound from the same prompt, audio (Replicate, PiAPI) [E].

A weak prompt and a stronger one (my own examples, not tested against a specific model):

Weak: A potter making a bowl, cinematic.

Stronger: Medium shot, eye level, slow dolly-in. A potter in a cluttered studio lifts a freshly thrown bowl off the wheel and turns it toward the window. Late-afternoon sun from the left, dust in the air, muted earth tones, 35mm documentary look. Sound: the wheel winding down, a distant tram, no music.

Habits that follow from the evidence and from common sense:

One action per clip. Clips are 8 to 10 seconds on mainstream models. Stack three actions, and you’re asking for a montage.
Name the light source. One practitioner guide argues real sources (“window light from the left”) beat mood adjectives like “dramatic” (Atlabs) [E].
Write the sound. If the model produces audio, leaving it blank leaves it to chance.
Anchor with images. Reference photos and first/last-frame control are standard on newer models, and they’re your best defense against a character changing faces between shots.
Keep a log. Model, version, date, prompt, settings. When Veo 3.1 left the API, anyone with that log could rebuild their look on another model far faster than anyone without it.

Music

Two years of lawsuits turned into licensing, mostly

The major labels sued Suno and Udio in June 2024. Then the posture changed. Universal settled with Udio on 29 October 2025 and announced a licensed platform planned for 2026 (MBW). Warner settled with Udio in mid-November, then became the first major to settle with Suno, which also took over Warner’s Songkick platform (MBW) [R].

What those deals meant for users is the part worth knowing. After the Universal settlement, Udio moved to a “walled garden”: generated music can be streamed on its platform but not downloaded or distributed elsewhere, and the company disabled downloads after a short window in November 2025 (Music in Africa, MBW) [R]. One comparison says Udio promises downloads will return with its licensed relaunch but has given no confirmed date (OrcaRouter) [E]. Practically: if you need files you can publish, Udio isn’t the tool right now.

Suno went the other way. On 9 September 2026, it released v6, built with content from Warner, BMG and Believe. Its chief product officer told Bloomberg that Suno will pay labels and publishers for the music it trains on, with labels deciding how to split royalties; financial terms weren’t disclosed. The model can edit parts of an existing song, build mashups, and generate from videos, photos, or journal entries (Gizmodo, Ece Yildirim, Suno) [R].

Complications sit right beside that good news, from the same Gizmodo report: Sony and Universal’s suit against Suno continues, Suno has admitted using YouTube audio for training, and a breach reported by 404 Media pointed to thousands of hours of scraped music [R]. And the American Federation of Musicians is suing Universal and Warner, arguing their AI licensing deals triggered a “new use” clause in its agreement and that session musicians should be paid; the majors have moved to dismiss, and I found no ruling (MBW) [R].

The music tools, today

Tool	Status	What to know	Evidence
Suno v6	Released 9 Sep 2026	Licensed-catalog model; Sony and Universal suit continues	[R]
Udio	Walled garden since late 2025	No downloads; relaunch date unconfirmed	[R] / [E]
ElevenLabs Music	Available	Vendor says trained on licensed data via Merlin and Kobalt deals; “cleared for commercial use” is the vendor’s wording, and one summary notes self-serve plans exclude some film, TV, and game uses	[E] (Invideo summary)
Google Lyria 3	Launched Feb 2026 with vocals; prompt guide in Google’s docs	Outputs carry SynthID watermarking per one source	[V] guide exists; [E] the rest
Stable Audio 2.5	Launched Sep 2025	Marketed on licensed training data; no vocals	[E]
Spotify AI remix/covers add-on	Not launched	UMG, Merlin, Kobalt and Believe have signed; launch reportedly planned for Q1 2027 for $4–5 a month under consideration; Sony and Warner reportedly still negotiating	[R] (MBW on the FT report)

The streaming data is the most honest number in this piece

Deezer publishes its own detection data, which is unusual. Fully AI-generated tracks went from about 10% of daily uploads in January 2025 to more than half by June 2026, roughly 90,000 tracks a day (Deezer newsroom, TechCrunch) [V].

The other half of Deezer’s data: fully AI-generated music is only 1–3% of listening, and the company says up to 85% of the streams on those tracks were fraudulent in 2025, against 8% across its whole catalog. Deezer demonetizes those streams, keeps AI tracks out of recommendations and editorial playlists, and will delete AI tracks tied to fraud or unplayed for six months [V].

Put plainly, the flood is mostly uploads, not listeners. If anyone pitches “upload AI songs, collect royalties” as a plan, this is the dataset to read first.

Where AI music is a defensible tool

From these cases: drafting a demo to show a collaborator, scoring a rough cut, producing a stem you’ll rework, background audio for your own video. Dr. Dre has said he uses AI in production (Gizmodo, August) [R], a reminder that professionals mostly use it as a sketchpad.

Prompting music (working practice, not documentation)

I couldn’t find a vendor-neutral, authoritative prompting guide for Suno, so what follows is practice. Google does publish a Lyria prompt guide in its documentation; check that and Suno’s current help pages before relying on any syntax.

Weak: A sad song about leaving home.

Stronger: Slow-tempo indie folk, around 72 BPM. Fingerpicked acoustic guitar, upright bass, soft brushed drums. Warm, close-mic’d female vocal, breathy, no vibrato. The verse builds quietly, the chorus opens up with harmonies, the song ends on a single guitar note. Lyrics: a kitchen at 5 a.m., a packed car, an unsent text.

Describe, don’t name. Skip living artists’ names. It invites a rights headache and tells the model who, not what.
Three instruments, not eight. Dense arrangements in a prompt tend to turn to mush.
Write the lyrics yourself if the words matter. Models are better at sound than at saying something specific.
Screenshot your plan’s commercial terms at generation time. With models being retired and replaced under licensing deals, what you were allowed to do last year may not be what you’re allowed to do now.

3D

What the evidence actually shows

I went looking for an audited case of AI-generated 3D assets shipping in a commercial title with before-and-after numbers. There isn’t one I could find. There are, however, three useful smaller pieces.

A studio’s own account. Paul Stashkevich, lead 3D artist at RetroStyle Games, wrote on 29 October that a mesh from an AI generator doesn’t go straight into a game: “a very good start,” in his words, but in their production the answer is no. Tripo provides the hi-poly mesh and the team does the rest, and in their game Last Pirate the tool is used at the prototype stage while the shipped assets are finished by hand. He notes that the numbers are rough estimates from daily production, not lab measurements (RetroStyle Games) [R].

A small comparative study. A practice-led study published in 2026 built the same stylized fantasy tavern two ways, by hand in Blender and with an AI-assisted pipeline using Meshy 6 and Hunyuan3D. The AI-assisted version took 238 minutes against 716 for the human-authored scene, but came with dense triangulated geometry, fragmented UV maps, inconsistent prompt adherence, clipping during placement, and lost texture quality when decimated (Zenodo record, journal page) [E]. It’s one scene, one researcher, in a small journal, so I’d treat it as a vivid example, not a benchmark.

Compare that with a vendor testimonial on an aggregator’s page, where one studio reports going from 14 hours to 1.5 hours per asset, a ninefold gain (3D AI Studio) [E]. The independent-ish study says about 3x with cleanup debt; the page that sells the tool says about 9x. Neither is wrong about their own scene. That gap is the lesson.

A sentiment check. The 2026 GDC State of the Game Industry survey (more than 2,300 respondents) found 52% of game professionals think generative AI is hurting the industry, up from 30% a year earlier and 18% the year before (GDC) [V]. About 36% use gen AI tools themselves, and only 7% said the impact is positive (GamesBeat) [R].

For a game with a community that cares, how you disclose and use AI 3D matters as much as the tool. That’s about reception, not capability.

The 3D tools, today

The field narrowed this year. One review says Luma’s Genie was retired on 1 January, ByteDance’s Seed3D 2.0 has no self-serve product, and Google’s acquisition of CSM left that roadmap uncertain, so for most studios the practical choice is now Meshy or Tripo (Frame Sixty) [E]. Frame Sixty is an AR/VR development studio that sells services, so weigh it as an informed practitioner, not a neutral lab.

Tool	Positioned as	What to know	Evidence
Meshy 7.1	Balanced all-rounder; rigging, 600+ animations, low-poly mode	Released 10 Sep 2026; reported 4096³ voxel “Ultra” mode accepts single-image input only; a Meshy Edit tool (16 Oct) edits one region but returns it untextured	[E]
Tripo (3.1 / Smart Mesh P1)	Fast iteration, hard-surface topology, auto-rigging	Reviewers cite ~10 s to a mesh; Smart Mesh P1.0 announced in March by the vendor	[E]
Rodin (Hyper3D)	Film-grade detail, hard-surface	Highest price tier among the hosted options	[E]
Hunyuan3D (Tencent)	Open weights, PBR	Read the license; one report lists regional limits on self-serve use	[E]
TRELLIS.2 (Microsoft)	Open image-to-3D model	4B parameters, up to 1536³, PBR attributes incl. transparency	[V] (project page, model card, paper)

Two flags. In a blind test of 605 senior 3D artists at NetEase and Tencent, Meshy 6 reportedly beat Tripo 3.1 by 63.8% to 36.2%; that result comes from a vendor-adjacent review, so call it suggestive [E]. And TRELLIS.2’s project page carries a disclaimer limiting its materials to academic and research use, while third-party coverage reports the model weights as MIT-licensed. A page disclaimer and a model license are different documents, so read the actual license file in the repository before shipping anything. One more licensing trap: one review says Meshy’s free-plan output is licensed CC BY 4.0, which requires attribution, while paid plans grant commercial rights [E]. Check the plan you’re on.

Prompting 3D: it’s mostly not prompting

Most strong results start from an image, not a sentence. TRELLIS.2 is image-to-3D by design, Meshy’s recommended game workflow starts from concept art, and RetroStyle begins from concept views made in an image model. The picture you feed in matters more than the words.

A reference-image brief that suits this workflow (my own, untested):

Low-poly stylized wooden treasure chest, lid closed, single object, three-quarter view, neutral gray background, flat even lighting, no ground shadow, no props.

Then the checks nobody puts in the demo video:

Set a triangle budget before you generate. Raw outputs can be enormous; the Meshy 7.1 review cites meshes of up to 80 million triangles before simplification.
Retopologize and bake. Every practitioner source I read treats this as required.
Rig it once as a test. Topology that looks fine statically can tear when it bends.
Check scale and materials in your actual engine. Meshy corrected a unit-scale bug in its USDZ exports on 12 October, so assets exported before then may need a scale check (Frame Sixty) [E].

Prompt helpers that earn their place

You asked about prompt tools, not just generators. These are the helpers with a documented function, rather than the “mega prompt packs” that fill search results:

Helper	What it does	Evidence
Veo/Omni prompt rewriter (Google)	Expands short prompts with scene, camera and sound detail; documented for Veo on Vertex AI	[V]
“Expert prompter” meta-prompt	Show a text model a reference image and ask it to write a comprehensive video prompt, as Google suggests	[V]
Gemini Omni multi-turn editing	Revise a generated video by conversation instead of re-prompting from scratch	[V] Google’s page
Suno v6 editing and mashups	Edit part of a song or build from a photo or journal entry instead of starting over	[R]
Meshy 3D Agent	Chat-based loop that brainstorms, gathers references and makes concept images, with per-step credit approval; on the free plan since September	[E]
Your own prompt log	Model, version, date, settings, what you kept and why	Your own records

The last row is the least glamorous and probably the most valuable.

Rights and records

The most useful legal fact for anyone prompting is that prompts alone don’t make you the author. The US Copyright Office’s January 2025 report on copyrightability concluded that, with currently available technology, prompts alone don’t give users enough control over the expressive result to count as authors (US Copyright Office, Part 2) [V]. On 2 March 2026, the US Supreme Court declined to hear Thaler v. Perlmutter, leaving in place the appeals court ruling that copyright requires human authorship; that case involved a work claimed to be made autonomously by AI, so it didn’t settle how much human input is enough (Mayer Brown) [R]. Law firms advising companies on this say the practical step is to preserve evidence of the human contribution to the final work (National Law Review) [R].

What that means for your workflow, as a practical reading rather than legal advice: the parts of a project where you made the expressive choices (editing, arrangement, hand-finished assets, original lyrics) are the parts you can most plausibly protect, and the raw output is the least protectable. Keep the log. And since this varies by country, ask a lawyer if the asset is valuable.

A one-hour test you can run

I couldn’t run benchmarks for this piece, so here is the smallest test that would give you real data, and that nobody else has for your material.

Pick one subject you actually need (a shot, a 30-second cue, a prop).
Write one prompt using the structure above. Don’t tune it per tool.
Run it on two or three tools at their cheap tier, ten generations each.
Score every output yes/no on one question: would I use this?
Compute keeper rate and cost per keeper (price per generation ÷ keeper rate).
Do the finishing step (trim, remaster, retopologize) on your best result and time it.

Tool	Generations	Keepers	Keeper rate	Price per generation	Cost per keeper	Finishing time

If you publish the filled table with your prompts, you have original data this field badly lacks, and a reason for others to cite you.

Who should start where

My own judgment, not a tested recommendation.

If you are…	Start with	Skip for now
A solo creator making short social video	Gemini Omni or Kling, drafts at the cheap tier, 8–10 second clips	Long single-take prompts
An indie developer on a tight art budget	Image-to-3D (Tripo, Meshy, or TRELLIS.2 if you have a GPU) plus a retopology pass	Anything player-facing without a disclosure plan
A musician or producer	Suno for sketches and stems you’ll rework	Uploading raw output at volume
A marketer	Runway or Omni for concepts and animatics	Final-ad realism without legal review of likeness and IP

FAQ

What is the best AI video generator in November 2026? There isn’t one that holds across jobs. Google’s Gemini Omni replaces Veo in the Gemini app and has the most clearly documented status; Runway is positioned for directed shots; Kling is widely recommended on value. Sora has been discontinued.

Is Veo still available? The Veo 3.1 preview models in the Gemini API were scheduled to shut down on 22 October 2026, with Gemini Omni 1.1 Flash as the named replacement. Third-party trackers report Veo 3.1 continues on Google’s enterprise platform for now; confirm on Google’s pages.

Can I use Udio songs commercially? Udio has operated as a walled garden since late 2025, with downloads disabled and no confirmed date for their return, so you can’t export new tracks for distribution. Check its current terms.

Is AI music a viable way to earn royalties? Deezer’s data says fully AI-generated tracks were over half of daily uploads in June 2026 but only 1–3% of streams, and up to 85% of those streams were fraudulent and demonetized in 2025. That’s evidence against upload-at-volume as a plan.

Can I copyright something I made with AI prompts? In the US, the Copyright Office says prompts alone ordinarily don’t make you the author, and courts have required human authorship. Human-made contributions to the final work are the part most likely to be protected. This isn’t legal advice.

Are AI-generated 3D models ready for games? Treat them as drafts. Practitioners and the one comparative study I found describe dense geometry, fragmented UVs, and cleanup work, and a studio using Tripo says its shipped assets are finished by hand.

What I could not verify

Any hands-on benchmark of my own. I ran none.
Omni 1.1 Flash prices other than the 720p rate, which I confirmed only through a Google quote; the 360p, 1080p, and 4K rates come from press reporting.
Whether Veo 3.1 stays on Google’s enterprise platform past 17 November.
Whether Udio’s licensed relaunch has a date, and whether downloads will return.
Licensing terms for Hunyuan3D and for any plan other than those I cite.
ElevenLabs’ and Stable Audio’s training-data claims beyond what the vendors state.
The result of the musicians’ union suit and of Sony and Universal’s suit against Suno.
Any audited production numbers for AI-generated 3D assets in a shipped title.
Whether Veo’s prompt-rewriter behavior carries over unchanged to Omni.

One constraint this article can’t resolve

The licensing deals that make Suno v6 and Udio’s relaunch look legitimate were signed by labels. Whether labels can license their artists’ and session musicians’ work to AI companies without separate consent or pay is exactly what the musicians’ union lawsuit asks a court to decide, and I found no ruling. If it goes the union’s way, the footing under every licensed AI music product shifts, and nobody building on these tools today gets a vote.

Sources

Video

Music

3D and games

Rights

US Copyright Office: AI report, Part 2 · Mayer Brown on Thaler cert denial · National Law Review

Beyond Text & Images: The Best AI Prompt Tools for Video, Music, and 3D (What Held Up)

The short answer

Eight cases, and how far to trust them

Tools are rented