At Overchat AI, we analyzed anonymized usage data to answer a question: when given equal access to all models, which do people pick? The research below is based on our data from Jul 9, 2026 through Aug 23, 2026 that covers 1.14 million users, 1.28 million sessions, and 28.2 million AI interactions
TALLINN, Estonia, Aug. 28, 2026 /PRNewswire-PRWeb/ — Visual AI now accounts for 70% of deliberate model choices
Overchat AI users can switch between text, video, and image-generation models within a single AI session. Across the 28.2 million AI interactions analyzed during the period, Overchat AI found that when people deliberately switched to another model mid-session, they chose an image-generation model in 53.7% of cases and a video-generation model in 16.5% of cases. People chose another text model in 29.8% of cases. Together, media-generation models were chosen in an overwhelming 70.2% of cases.
Category – Share
Image – 53.7%
Video – 16.5%
Text – 29.8%
In 92% of cases, AI sessions began with a text model. However, people who did switch typically moved into media-generation workflows and, in 72% of cases, did not return to a text model before ending the session. This suggests that in a clear majority of these situations, people preferred to continue creating media rather than return to chatting or asking questions.
Most popular AI models
The table below summarizes the models people switched to most often. In other words, the sessions did not begin with these models; users selected them later in the conversation:
Model family – Share of deliberate choices
Seedream 5 Pro – 16.9%
Grok Imagine 1.5 – 15.9%
Nano Banana 2 – 12.6%
Kling V3 – 8.9%
Gemini 3.7 Flash – 8.2%
GPT-5.6 – 7.5%
Claude Opus 4.8 – 6.6%
Other models – 23.4%
It’s interesting to see that Seedream 5 Pro led deliberate model choices despite being a lesser-known model. Seedream is an advanced family of image-generation and editing models developed by ByteDance, known for precise image-editing capabilities, rich detail, and fast generation times. This could suggest that when people choose media models, they are less interested in who made the model and more focused on quality and user experience.
It’s also interesting that, among text-based models, Gemini overtook both GPT-5.6 and Claude Opus 4.8. Overchat AI observed a recurring pattern in which a chat would begin with GPT-5.6 before the user switched to Gemini. By comparison, users switched back to the GPT family less often.
For initial model choices, sessions most frequently began with Nano Banana 2 or GPT Image 2 for images and GPT-5.6 for text. This suggests that people initially gravitated toward familiar names but quickly moved on to explore other options—and often stayed with them.
Multimodal AI use is becoming mainstream
About 84% of active users worked across more than one modality during the research period, combining text with image or video generation rather than staying within a single workflow.
While 92% of AI sessions began with a text model, users who moved into media generation rarely returned to text during the same session. In 72% of those cases, the session ended in the image or video workflow.
Mobile users gravitate even more strongly toward visual AI
Mobile accounted for 76.4% of users and 75.3% of sessions during the period. Visual models accounted for 72.1% of deliberate choices on mobile, compared with 64.7% on desktop.
Device – Visual share of model choices
Mobile – 72.1%
Tablet – 69.4%
Desktop – 64.7%
Smartphones are becoming an AI creation environment, and image and video generation appear particularly well suited to mobile-first behavior.
The first model switch reveals the user’s actual task
Although 92% of sessions began with a text model, the first transition most often took users into image generation. Moving from one text model to another accounted for only 21% of first switches.
First transition – Share
Text to image – 54%
Text to video – 17%
Text to another text model – 21%
Media to another media model – 5%
Media to text – 3%
People often changed models more than once
Among sessions that involved more than one model, users switched models 1.8 times and interacted with 2.6 different models on average. Around 41% of multi-model sessions included at least two separate model changes.
Model changes – Share of multi-model sessions
1 – 59%
2 – 21%
3 – 10%
4 – 5%
5 – 3%
>5 – 2%
Most model switching happens early in the session
Overchat AI found that people usually changed models before becoming deeply invested in a conversation. In 58% of multi-model sessions, the first switch happened within the first third of the session. Only 14% of first switches occurred near the end.
First model switch – Share
First third of the session – 58%
Middle third – 28%
Final third – 14%
Most users finish with a different model than they started with
Only 15% of multi-model sessions ended with the same model that started the conversation. In 57% of cases, the session ended in a different AI modality altogether.
How the session ended – Share
In a different modality – 57%
With another model in the same modality – 28%
With the original model – 15%
Multimodal sessions are substantially deeper
Sessions became substantially longer when users combined different AI modalities. Text-only sessions averaged 14.2 interactions, compared with 25.7 interactions for sessions that combined text and image generation.
Session workflow – Average AI interactions
Text only – 14.2
Text and image – 25.7
Text and video – 31.8
Text, image and video – 42.4
The deepest sessions were those that included text, image, and video models, suggesting that people are using several models to develop, revise, and transform an idea within the same session.
Methodology
The research covers anonymized, aggregate activity on Overchat AI from Jul 9, 2026 through Aug 23, 2026. It is based on product behavior patterns. “Deliberate model choices” refers to occasions when users actively selected a different model after starting a session with another one. The multimodal-use, switch-frequency, transition, session-depth, and device figures were derived from aggregate session behavior. No individual accounts, prompts, uploads, or generated content were reviewed.
Key findings
Visual models accounted for 70.2% of deliberate model choices.About 84% of active users worked across more than one AI modality.While 92% of sessions began with text, 72% of sessions that moved into media generation ended there.Multi-model sessions included 1.8 model switches and 2.6 different models on average.Users averaged about 22 AI interactions per session.Seedream 5 Pro was the most popular AI model among deliberate switches.About 76% of users accessed AI on their phones.In 58% of multi-model sessions, the first switch happened within the first third of the session.Text-to-image transitions accounted for 54% of first model switches.Only 15% of multi-model sessions ended with the model that started the conversation.Sessions combining text, image, and video averaged 42.4 AI interactions, compared with 14.2 for text-only sessions.Visual models accounted for 72.1% of deliberate choices on mobile and 64.7% on desktop.
About Overchat AI
Overchat AI is an all-in-one AI app that gives users access to leading text, image, video and audio models from OpenAI, Anthropic, Google, xAI, ByteDance and other providers in one place. With one account, people can compare models, switch between them and use more than 150 purpose-built AI tools for writing, research, image creation, video generation and everyday tasks.
Learn more at overchat.ai.
Media Contact
Ekaterina Hohlova, Overchat AI, 372 53236240, ehohlova@overchat.ai, https://overchat.ai/
View original content to download multimedia:https://www.prweb.com/releases/new-data-from-28-2-million-ai-interactions-reveals-how-people-use-ai-302861717.html
SOURCE Overchat AI
