The two breakouts inside the numbers
Two shifts carry most of the change. In image, Alibaba's Qwen Image doubled its share from 12.3% to 26.0%, nearly catching Google's Nano Banana (26.1%) and closing in on ByteDance's Seedream (33.5%). In the final two weeks it became the single most-used model family across all categories at 17.2%, ahead of GPT text models (10.9%) and Gemini (10.8%). In video, ByteDance's Seedance jumped from 10.8% to 25.2% on the strength of Seedance 2.5, while xAI's Grok Imagine fell from 39.6% to 30.6%. Kuaishou's Kling held 21.9%.
The price lines are worth a second look. Seedance 2.5 costs roughly three times as much per video as Grok Imagine 1.5 on the platform, and it still gained. Qwen Image costs the same as OpenAI's GPT Image and both sit in the free tier, yet Qwen was used about four times as often during the final two weeks. Price explains neither outcome. People paid more for the video model they wanted and picked the free image model they preferred.
Who changes their mind
The new-user numbers read like a story about defaults. OpenAI and Anthropic account for 23.8% of new users' model usage but 19.8% among returning users. Google rises from 20.9% to 28.0%; Chinese providers from 36.3% to 39.8%. Together Google and Chinese models take 67.8% of returning users' usage. Familiarity picks the first model; experimentation picks the second.
Geography bends the split further. The Chinese-model share reaches 43.2% in the US, 44.6% in the UK, 46.4% in Australia and 52.0% in Spain. Japan is its own planet: Gemini accounts for 74.4% of usage there. The US is the only major market where video is the leading category (36.8%); Brazil is the most text-focused (49.3%). Mobile, at about 74% of sessions, splits almost evenly across text, image and video; desktop is 51.3% text, led by GPT, Gemini and Claude.
The footnotes, read properly
This is one app's data, published by the app itself. Overchat AI's users chose to install a multi-model playground, which skews experimental; it is not a representative sample of all AI users. And the methodology counts each user once per model per period, so one curious dabbler and a heavy daily user carry the same weight. Neither caveat invalidates the direction: the shift is large, consistent across categories, and visible in both new and returning cohorts.
One more number to hold against the hype. Google still leads overall usage at 23.1% by developer, and OpenAI posted the biggest jump of any lab, 8.8% to 13.6%, on GPT-6 and GPT Image 2.5. The American labs are not losing. They are losing specifically the visual category, where the fight is over outputs people keep and share, and where Chinese labs have shipped faster, longer, and cheaper than anyone else for two years.
The practical takeaway
If you build or buy creative tooling, the text-model leaderboard is the wrong map. The models people actually pick for images and video look nothing like it, and the gap keeps widening. Anyone still choosing a video stack from brand familiarity alone is reading last year's chart.
Sources
- [1] Overchat AI via EINPresswire — “People Choose Chinese AI Models for Image and Video Creation More Than Western Models, New Data Shows” (October 8, 2026)Read source
- [2] Overchat AI Q3 2026 report (July 17, 2026)Read source