Seedance 2.5 Flash now 35% off — ends Nov 30

AI creation is becoming everyday work. Here's what our data shows.

In January, most companies on Atlas Cloud were still testing generative AI: a few calls, a few comparisons, then back to the old workflow. Only 18% of business accounts were in...

AI creation is becoming everyday work. Here's what our data shows.

In January, most companies on Atlas Cloud were still testing generative AI: a few calls, a few comparisons, then back to the old workflow. Only 18% of business accounts were in steady use. By October, 67% were.

That's the story of our year. We think 2026 is when AI creation stopped being a side experiment and started becoming everyday work, and we expect 2027 to make it routine.

We published the numbers behind that view today in The Model Mix 2026, our first report on how builders use generative AI models: ten months of usage on Atlas Cloud, January to October 2026, across 400+ image, video, audio, 3D and language models. Here's what we saw, and what we make of it.

"In January, most companies on Atlas Cloud were testing AI. By October, two-thirds were in steady use."

Companies stopped testing and started building

The number of business accounts in steady use on Atlas Cloud grew about five and a half times from January to October. Steady use means calls in five or more separate model-hours in a month: work that comes back, not a single trial.

Growth was fastest from April to July. A new generation of video models started arriving in April, with Dreamina Seedance 2.0 and Wan 2.7, followed by Dreamina Seedance 2.5 and Wan 3.0 in August. Usage data can't prove cause and effect, but our view is clear: video crossed a quality bar this year, and teams that had been waiting for output they could put in front of a client started building. That's also what the studios and brands we work with tell us. Versatile Media's MOKE, for one, runs its film pipeline on Atlas Cloud, from script to final cut.

These companies work like professionals. Business accounts in steady use ran 10.5 model endpoints on average in October, against 7.6 for other accounts in steady use, and kept about one in four of the models they tried. Serious teams treat models like a bench of specialists: test several side by side, keep the few that earn a place.

A new wave of creators showed up

Companies grew, and individuals grew faster. Active accounts signed up with a personal email more than doubled on Atlas Cloud this year, most of it between May and August. We read these as mostly creators, prosumers and developers, though some are people using a personal address at work.

We think this group is underrated. Tools matured, workflows got easier to learn and prices came down, and that put capable models within reach of one person with an idea: a short film, a social channel, an AI influencer account. The video settings we saw across Atlas in late October fit that picture, with vertical 9:16 edging out 16:9, 15- and 5-second clips among the most common lengths, and most clips carrying sound.

These people are teaching themselves AI production on their own time. Brands and studios will be hiring that skill in 2027.

"Individual creators more than doubled on Atlas Cloud this year. They're the AI production talent brands will hire next."

The prompt box is no longer where the work starts

In October, 88% of image accounts on Atlas used editing, up from 54% in January, and 96% of video accounts animated an existing image.

We think the prompt-box era is ending for production work. A starting image fixes the product, the brand and the composition in a way a text prompt can't, so we believe teams get usable output in fewer tries. The pattern fits a simple workflow: product images in, versions out. Place the product in new scenes, then animate them.

That changes what a creative team's most valuable asset is. It isn't the cleverest prompt. It's the library of approved product and brand images that every version starts from. Brands and retailers with large catalogs start 2027 ahead, and agencies that build those image-to-version workflows for their clients will matter more, not less.

"A creative team's most valuable AI asset isn't its prompts. It's its library of approved images."

LLMs and media are becoming one stack

81% of the accounts that used a language model on Atlas in October also made media, up from 33% in January, and seven in ten of them started on Atlas with media. The median account using both made 28 text calls in October. We infer the language model is a helper, writing scripts, captions, ad copy and prompts.

The same pattern runs across every kind of media. 64% of accounts in steady use worked in two or more modalities in October, up from 35% in January. Audio went from under 1% of accounts to 11% after we added most audio models in June, and builders added it to image and video work rather than buying it on its own.

Our view: the line between "LLM companies" and "media companies" no longer describes how people build. The next products will write, see, speak and render in one flow, and the teams building them want every one of those models in one place.

"The line between LLM companies and media companies is disappearing. 81% of LLM accounts on Atlas Cloud also make media."

No model stays on top for long, so don't build on just one

The model used by the most accounts changed in image, video, text-to-speech and 3D on Atlas this year. The image lead changed hands three times. In 3D, Tripo went from 5% to 49% of 3D accounts in three months.

Builders handled that by keeping their options open. The average account used 6.2 model endpoints in October, up from 3.4 in January, and the share relying on a single model fell from 42% to 27%. About 60% of accounts that changed their main model family still used the old one the next month: they add the new leader before they drop the old.

We think that's exactly right. Betting a product on one model is a bet that it stays best, and in four of the media tasks we track, the leader changed within the year. Build so you can switch: route each task to the model that's best for it today, and test every new release on your own work. Yapper, for example, adds frontier models through one integration, with no new vendors.

Trying widely also goes with staying. Of accounts new to Atlas from November 2025 to July 2026, 44% of those that tried five or more models in their first week were still active three months later, against 8% of those that tried one. Busier accounts both try more and stay longer, so that's an association, not proof, but it points the same way.

"No model stays on top for long, so no serious team should build on just one."

The biggest launches bring in new builders

The fastest image and video launches reached a quarter of their modality's launch-month accounts within nine days of first use on Atlas.

What struck us more is who they reached. Of the accounts that made Dreamina Seedance 2.5 their main video model in its first two months, 52% had not made video on Atlas before. For Dola Seedream 5.0 Pro, 63% were new to image on Atlas. After a heavy price cut with a campaign on a budget video tier, the tier drew 2.2 times as many first-time users in four weeks as in the four weeks before, and 67% of them were new to Atlas.

Our read: on Atlas, the biggest launches grow the number of people building more than they reshuffle the people already here. A real step up in quality, or down in price, brings in people who weren't making video or images on Atlas before.

What we expect in 2027

  • More production starts from an existing image. Marketing and e-commerce teams with large libraries of approved images move fastest.
  • Video becomes everyday work for brands. Companies keep moving from testing to steady use, and video is the format that pulls them in.
  • Language models get bought as part of media work. Media teams add LLMs as helpers for scripts, captions and ad copy.
  • Builders keep several models in use. Switching becomes a design choice, not an emergency, and every new release gets tested on real work within days.
  • 3D goes deeper before it goes wide. The average 3D account made about 4.7 times as many successful requests in October as in June. We're watching 3D rather than calling it, but the builders who use it are using it hard.

What this data is and isn't

This is one platform's data, and every figure is on Atlas Cloud, not the whole market. Atlas Cloud is an API gateway serving image, video, audio, 3D and language models through one API key. Its accounts lean toward media builders, and companies committed to a single model provider mostly call that provider directly. Each account counts once, whatever it spends. Account totals are not published; findings are shares, averages and growth over the year. Atlas runs promotions and price cuts on the models it serves, which shapes some of these numbers.

Read the report

The full report covers every modality, which models lead and why, how price and benchmarks relate to usage, and what the data means for brands, studios, creators and model makers. It's free, with no form: Read the report. Want the next one? Get the next edition.

Ready to build with these models? Explore the models on Atlas Cloud: video, image, audio, 3D and language models through one API.

If your numbers look different, tell us. We'd like to hear how you see it.

Latest Models

One API for All Media AI.

Explore all models