Atlas Cloud Research. First edition, November 2026
The Model Mix 2026
Which image, video, audio and language models builders use, how they use them, and where the market goes in 2027. Ten months of usage on Atlas Cloud, which serves 400+ models.
By Colin Mo, Head of Product Marketing, Atlas Cloud (LinkedIn)
- ✓One platform sees image, video, audio, 3D and language models on the same account.
- ✓Ten months of real usage, January to October 2026, not a survey.
- ✓Compared with fal and OpenRouter where they measure the same thing.

Summary
What changed in 2026
Most generative media work on Atlas Cloud now starts from an image, and the people doing it changed in 2026: companies moved from testing to steady use, and individual builders more than doubled.
88%
of image accounts used editing in October, up from 54% in January
67%
of business accounts were in steady use in October, up from 18% in January
5.5x
growth in business accounts in steady use, January to October
2x+
personal-email accounts more than doubled, most of it from May to August
- 88% of image accounts used editing in October, up from 54% in January, and 96% of video accounts animated an image.
- Our reading: a starting image pins down the product, the brand and the composition in a way a prompt alone does not, so teams get usable output in fewer tries.
- Companies moved from testing to steady use. In January, 18% of business accounts were in steady use; in October, 67% were. They run more models than other accounts (10.5 endpoints against 7.6) and keep about one in four of the models they try.
- Our reading: brands and e-commerce teams put product images in and get versions out, studios and short-drama producers drive video volume, and software teams start on language models and add media.
- Business use deepened fastest from April. The number of business accounts in steady use grew about five and a half times from January to October, fastest from April to July, alongside new video models: Dreamina Seedance 2.0 and Wan 2.7 in April, then Dreamina Seedance 2.5 and Wan 3.0 in August, which now sit near the top of the video boards.
- Our reading, from the studios and brands we work with: better video models likely helped, as filmmaking, studio, advertising and other media-based work grows now that models clear the bar those teams need. The usage data alone can't show the cause.
- Individual builders grew faster than companies. Personal-email accounts more than doubled, most of it between May and August; business accounts grew about half.
- Our reading: maturing tools and falling prices put capable models within reach of creators, prosumers and developers, who experiment on their own projects, some of them building influencer accounts.
- Builders work across media: 64% of accounts in steady use touched two or more modalities in October, up from 35% in January, 81% of language-model accounts also make media, and audio went from almost nothing to one account in ten.
- The lead changed hands in image, video, text-to-speech and 3D, and builders add the new leader before they drop the old one.
- Usage doesn't simply follow benchmarks or price: on none of six leaderboards was the top-rated model we could match the most used, and price barely related to share, apart from one heavy cut.
Get the next edition
We'll send the next Model Mix when it's out, plus data updates in between. Nothing else.
Method
About this report
What Atlas Cloud sees
Atlas Cloud is an API gateway that sells models as a service: it hosts some models itself and passes calls through to the model's provider for others. Because one account reaches image, video, audio, 3D and language models, we can see how the same accounts combine them. It is a selected slice, mostly media builders and buyers of open-weight models on price; companies committed to OpenAI, Anthropic or Google mostly call those providers directly. We compare with fal, a media-first inference platform, and OpenRouter, a language-model router, where they measure the same thing.
How we counted
- The unit is an account, not a customer or a person. An account is active in a month when it made at least one successful call to a generative model; Atlas's internal accounts, apps and utilities are left out.
- Every account counts once, whatever it spends. One company running production work can use more than thousands of individual accounts combined, so account shares can understate how much of the work businesses do.
- Accounts in steady use made calls in five or more separate model-hours in the month, and the main headlines were rerun on them.
- We give shares and correlations, never account counts, and no account-based figure rests on fewer than five accounts.
- The window is January to October 2026, and every figure is "on Atlas Cloud".
- Request-setting snapshots, the daily price series and the moderation-rejection figure come from model-level aggregates that can't exclude Atlas's internal accounts; every other figure does.
- Request-settings data (shapes, sound and references) covers October 19 to 31 only.
How we handled the data
- We used usage records, not content. Every Atlas usage figure comes from billing and task records (which model an account called, when, for what kind of task, with how many input files and which output settings) and two account attributes used to group accounts: whether the sign-up email is a company address, and whether it is a team account. Benchmark ratings and other platforms' figures come from the published sources we cite.
- We opened no customer content for this report. No prompts, uploaded images, generated images, video, audio or text were opened, sampled or used.
- We publish only aggregates and name no customer in the findings. Figures are aggregated across accounts and published as shares, with no account counts, and no account-based figure rests on fewer than five accounts. Request-setting figures are shares of successful tasks from October 19 to 31; that extract carries no account field, so we could not run the five-account check on them. No customer is named in the findings.
About Atlas Cloud
The AI Creation Engine for creators and developers.
Atlas Cloud helps teams create content and build AI-powered products with leading AI models. We bring video, image, audio, and language models together through one API, alongside no-code creation tools and enterprise support. With competitive pricing and enterprise-grade reliability, we serve studios, agencies, brands, and AI product teams.
- Model API: Leading models. One API. Connect leading models to your products and workflows, and try different models through one account and one API.
- Creation Tools: Create, compare, and iterate. Generate content, compare models, create marketing assets and manage your files with Generator, Model Explorer, Marketing Studio and Asset Library.
- Enterprise: Built for production. Enterprise technical support, volume pricing, and dedicated infrastructure and deployment options for supported models.
From Models to Creation.
Try the models in this report
Image, video, audio, 3D and language models through one Atlas Cloud API key and one bill.
Sources
- fal, The State of Generative Media, Volume 1, February 2026, and Volume 2, September 2026.
- OpenRouter, rankings (image, video and speech, week of October 26, 2026), The Token State of Play (September 22, 2026) and DeepSeek V4 Is Earning Agentic Token Share (June 30, 2026).
- Artificial Analysis leaderboards: text-to-image, image editing, text-to-video, image-to-video, text-to-speech and the Intelligence Index, monthly captures, 2026.
Journalists and analysts: email marketing@atlascloud.ai for chart files.
Data covers Atlas Cloud accounts, January to October 2026. Account totals are not published. Findings include shares, averages, correlations and the median of 28 text calls. Atlas Cloud serves nearly every model in this report.