The AI creative tool map turns four months old. Here's what changed since

On the night of Tuesday 26 May, Minh Do was still adding logos. "I was literally up late yesterday still adding companies to the list," the Machine Cinema cofounder wrote on X the next afternoon, when he published the AI Creative Market Map: more than 250 tools sorted into five phases, running from the first draft of a script to the streamer that pays for the finished film. A year earlier the same exercise had produced about 100 names. In the Substack essay that went out with it, he gave his own map a short shelf life and wrote that it "will probably be useless tomorrow."

We're four months late to it, and that turns out to be useful. Machine Cinema's map is still the clearest single picture of the tools behind AI film, ads and music video, and four months is long enough to check its calls against what actually shipped. Most of them held, and one got stronger. What follows is the map phase by phase for anyone starting out, what changed in each phase since May, and where to begin if you would rather not learn 250 tools.

How to read the map

The map runs left to right in five phases: Development, Production, Inference, Post-Production and Distribution. Inside them sit 19 categories by Machine Cinema's count, from storyboarding and voice to world models, upscaling and microdramas. Minh Do compiled it from the community's GenJams, the hackathon-style sessions where its members make work with new tools in public, and he runs Machine Cinema with the Los Angeles producer Fred Grinstein. The landing page is free, and the full-resolution image sits on the Substack with no email gate.

There are two things to know before you lean on the numbers. Many companies appear in several boxes because they keep expanding into each other's turf, so KREA turns up in at least six places and ElevenLabs in four. Minh's line for it is the best one in the essay: "Everyone wants to be everything. For lack of moat, become the ocean?" The growth curve is also Machine Cinema's own framing rather than a census. The four tools of 2023 were the ones on hand at the community's earliest GenJam (ChatGPT, Runway, DALL-E 3 and Midjourney), the 2025 edition had no Distribution phase at all, and companies are invited to add themselves through the website.

Machine Cinema also works with many of the companies it maps. Its partner page sells sponsored workshops and co-branded content and names Google, ByteDance, Luma, ElevenLabs, Runway, OpenAI and Midjourney as tool partners, while OpenArt, which sits in the map's aggregator box, presents the community's Director's Cup. That is normal for a community that runs on sponsored jams, and nothing we found suggests anyone paid for a place. Read it as one well-connected room's view of the market, which is exactly what makes it worth reading.

Development: the chatbot still writes the first draft

The first phase covers story development, storyboarding and previz, and production assistants. For a beginner it is the part of the map that has changed least. Writing still runs mostly through general-purpose chatbots, and the storyboard tools have drifted toward the aggregator model, bundling several image and video models behind one interface. If you already plan with a chatbot and a sketchbook, this phase is covered.

Production: the models still rule

Production is by far the biggest block, covering image and video models, voice and dubbing, lipsync and avatars, music, sound effects, 3D, world models and the aggregators that bundle them. Minh's read was blunt: "To some extent, we are all downstream of model providers." That has held. ByteDance's Seed team launched Seedance 2.5 on Friday 31 July with 30-second single takes, and it was running inside Runway by Friday 7 August (the map still shows Seedance 2.0). MiniMax shipped H3 the same day as Seedance 2.5, with native stereo sound and clips of up to 15 seconds at 2K, and Alibaba formally released Wan3.0 on Monday 24 August, hosted only, at 5 to 20 cents per second of output.

Kuaishou announced Kling 4.0 on Monday 28 September, and its real upgrade is control, with ten keyframes and a 15-item reference budget. Only the lighter Flash tier is live, at 720p for top-tier subscribers, and the full model is due in October with no date, price or API. Voice and music moved as well. ElevenLabs v4 turned the job from generating a read into directing a performance, and Suno's v6 arrived with three major labels behind it, though songs made on the free tier still cannot legally go under a commercial cut. Google's next Nano Banana has been spotted inside Flow but has not launched.

The company missing from the video box is OpenAI. Sora was on the 2025 map and is absent from this one, and OpenAI has since finished the job, shutting the Sora app on Sunday 26 April and discontinuing the API on Thursday 24 September with a one-line statement: "We've decided to discontinue Sora in the consumer app and API." Pika changed boxes too, from video generation to lipsync and avatars, after pivoting in April to what it calls "AI Selves: persistent, agentic beings." Runway's Project Continuum, a desktop where some windows are generated video, is the most ambitious agentic idea since the map, and for now it lives in a demo reel.

Inference: the reason there are so many aggregators

Inference is the plumbing: companies such as fal, Replicate, Baseten and Runware rent out the computing power to run models through an API. Most beginners will never touch them directly, but they explain why the rest of the map looks the way it does. Because anyone can wire a new model into an app within days, the aggregator box became the most crowded on the map, and new models keep turning up inside other companies' products almost as soon as they ship. Runway hosting ByteDance's Seedance is the cleanest example of a company choosing to become the ocean.

Post-production: the quiet box that got bought

The post-production phase covers upscaling, color grading, generative VFX, AI editing, and legal and likeness tools. Its upscaling box held Topaz Labs, Magnific, KREA and Adobe Firefly, so Adobe and Topaz sat side by side in one small square. On Thursday 25 June, four weeks after the map came out, Adobe announced it was buying Topaz, and the deal closed on Wednesday 23 September. Adobe's SEC filing puts the price at about $340 million, a figure neither announcement gave. Runway has since added Enhance Frame Rate, which conforms any shot to ten delivery frame rates, and Runway and Adobe have both put generation inside Premiere and After Effects.

Legal and likeness is the box worth reading even if you never buy a tool from it. Since May, a 31 August order from India's Registrar of Copyrights said an AI system cannot be an author but the human behind it could be, a federal judge enjoined the disclosure label in Montana's SB 25 against one plaintiff and his PAC, and Apple's VFX crew got an IATSE contract whose AI protections were imported from the 2024 Basic Agreement. Each one changes what you can sign, sell or call your own.

Distribution: the problem the map says matters most

Distribution is the newest phase on the map and the thinnest, with streamers and microdramas. Minh's essay is clear that this is where the real fight is: "At the end of the day, the most important problems in the space are distribution and monetization." The four months since have mostly proved him right by showing how messy the money still is. Beijing's Huairou district now covers 30% of compute for AI video, capped near $3,000 a title and paid after release, so the subsidy rewards hits rather than experiments. Valve lets you run YouTube ads on a Counter-Strike video but bars sales and licensing, and its rules never imagined an AI remake. YouTube, meanwhile, is filling up with fake trailers for films that have barely shot a frame.

The cost side is just as uneven. List prices make an AI short look cheap, but the real bill is set by how many attempts it takes to get one usable shot. If you want proof that the money sits outside Hollywood, look at WPP's Devon's Point, a 34,000 square foot production house where AI is the wiring rather than the product.

The six calls, scored

Machine Cinema pulled six arguments out of the map. Here is how each one looks on Monday 5 October.

• Models still rule the space: held, since Seedance 2.5, MiniMax H3 and Wan3.0 all shipped after the map and Kling 4.0 is next in line.
• The revenue is not where you think: held, and now sourced, because ElevenLabs' chief executive put enterprise at more than 55% of about $600 million in annual recurring revenue in September, by the company's own unaudited figures, up from about 40% a year earlier.
• For lack of moat, become the ocean: held, as Runway hosts ByteDance's models, and Adobe bought Topaz rather than out-building it.
• The aggregator category exploded: held, because inference made wiring up a new model a matter of days.
• The Chinese have arrived: this call is stronger than in May, with Seedance 2.5, Wan3.0, MiniMax H3, Kling 4.0 Flash and a city government paying compute bills after a hit.
• We have barely seen agentic: still true, since Continuum is a reel rather than a product, and Pika's agents are avatars rather than filmmakers.

Where a newcomer should start

In practice, for anyone making work this week, you don't need 250 tools. You need one model you know well, one place to run it, and a plan for where the work gets paid. Pick a video model and learn how it fails before you learn its tricks; an aggregator is a cheap way to compare a few before you commit. Learn one post tool properly, because upscaling and frame rate conversion are where a rough generation becomes a deliverable. Read the license on your plan before you post anything commercial, since the free tier and the paid tier of the same model often carry different rights.

Then learn in public, which is Machine Cinema's own answer to the map: its GenJams and community are where the question Minh closes on gets worked out shot by shot: "What does an AI creator look like tomorrow?" The map shows the money sitting downstream of the models and outside Hollywood, which is a good place for a newcomer to build a workflow.

Since the map: the dates

• Date: Sun 26 Apr 2026. What happened: OpenAI shuts the Sora web and mobile app
• Date: Wed 27 May 2026. What happened: Machine Cinema publishes the AI Creative Market Map, 250+ tools in five phases
• Date: Thu 25 Jun 2026. What happened: Adobe announces it is buying Topaz Labs
• Date: Fri 31 Jul 2026. What happened: ByteDance launches Seedance 2.5; MiniMax launches H3
• Date: Fri 7 Aug 2026. What happened: Seedance 2.5 goes live inside Runway
• Date: Mon 24 Aug 2026. What happened: Alibaba formally releases Wan3.0
• Date: Mon 31 Aug 2026. What happened: India's Registrar of Copyrights rules on DABUS and AI authorship
• Date: Wed 23 Sep 2026. What happened: Adobe closes the Topaz deal, about $340 million per its SEC filing
• Date: Thu 24 Sep 2026. What happened: OpenAI discontinues the Sora API
• Date: Mon 28 Sep 2026. What happened: Kling 4.0 announced; Flash tier opens to top-tier subscribers
• Date: October 2026. What happened: Full Kling 4.0 due, no date given

Our coverage, phase by phase

• Production, video models: thegenerated.com/stories/candy-seedance-25-eight-minutes-20260917 and thegenerated.com/stories/kling-4-0-control-terms-20260930
• Production, voice and music: thegenerated.com/stories/elevenlabs-v4-voice-direction-20260929 and thegenerated.com/stories/suno-v6-labels-commercial-rights-20260924
• Production, image and agentic: thegenerated.com/stories/nano-banana-2-1-google-flow-leak-20261005, thegenerated.com/stories/runway-project-continuum-generated-video-desktop-20261002 and thegenerated.com/stories/luma-hybrid-underwater-pool-20260922
• Post-production: thegenerated.com/stories/adobe-topaz-labs-acquisition-ai-upscaler-20260930, thegenerated.com/stories/runway-enhance-frame-rate-20260918 and thegenerated.com/stories/edit-suite-ai-operator-20260924
• Legal and likeness: thegenerated.com/stories/india-ai-copyright-ownership-dabus-20261001, thegenerated.com/stories/montana-sb25-blocked-20260921 and thegenerated.com/stories/apple-vfx-iatse-contract-ai-clause-20261001
• Distribution and cost: thegenerated.com/stories/china-ai-video-subsidies-pay-for-hits-20260928, thegenerated.com/stories/valve-video-policy-ai-counter-strike-fan-films-20261001, thegenerated.com/stories/raaka-fake-trailers-youtube-20260929, thegenerated.com/stories/ai-short-film-real-cost-20260923 and thegenerated.com/stories/wpp-devons-point-ai-production-hub-20260923

Sources and what is unconfirmed: Machine Cinema's map landing page, partner page and the full-resolution map image; Minh Do's Substack essay and launch post on X (27 May); Kling's release note and blog; MiniMax's blog; Alibaba Model Studio and Winbuzzer on Wan3.0; Runway's Seedance pages; Adobe's blog and its SEC filing as reported by TV Tech; PCMag and AI Primer on Sora; Pika's blog; TechCrunch's interview with ElevenLabs' chief executive as relayed by RuntimeWire. The 250+ and 19-category counts are Machine Cinema's, and we did not count the logos ourselves. We did not open the email-gated PDF; the six calls are checked against the landing page and the Substack, which carry the same arguments. ElevenLabs' revenue figures are company-reported. Map artwork: AI Creative Market Map, Q2 2026, Machine Cinema / Minh Do, described here and not reproduced; rights holders can request changes or removal.