Imagine a desktop racetrack. One car talks, another paints, and a third makes moving pictures. They all look fast. But before naming a winner, ask what the clock is measuring: website visits, votes on an output, or the time it takes to complete your own project?
An app such as ChatGPT is a service people use; a model is an AI system working underneath. An app can offer multiple models, and a model can power multiple services. Traffic to one site therefore cannot count worldwide use of one model.
The chat track: Website visits Similarweb’s worldwide AI Chatbots and Tools website ranking for September 2026, updated 1 October:
| Rank | Website |
|---|---|
| 1 | chatgpt.com |
| 2 | gemini.google.com |
| 3 | claude.ai |
| 4 | chat.deepseek.com |
| 5 | openai.com |
Source: https://www.similarweb.com/top-websites/ai-chatbots-and-tools/ This is a website ranking, not a top-five model ranking: openai.com is the company website. It excludes usage through mobile apps, APIs and other products and does not establish unique users, satisfaction or answer quality.
Rules and practical implications
A different view comes from a16z’s seventh consumer AI report, published 5 October 2026. It separates Similarweb website visits from Sensor Tower monthly mobile active users and adds US consumer card spending. In its US panel, 4.5% of eligible consumers had at least one paid personal ChatGPT, Gemini or Claude subscription in August 2026. This is not a global or enterprise adoption rate. https://a16z.com/100-gen-ai-apps-7/
The image track: Blind preferences Artificial Analysis text-to-image leaderboard, accessed 3 November 2026:
| Rank | Model / setting | Elo | Estimated rank range |
|---|---|---|---|
| 1 | GPT Image 2.5 Sunburst (max) | 1223 | 1 |
| 2 | GPT Image 2.5 Flare (max) | 1207 | 2 |
| 3 | GPT Image 2 (high) | 1189 | 3–4 |
| 4 | Nano Banana 2.1 | 1176 | 3–6 |
| 5 | Grok Imagine Image 2.0 | 1168 | 4–6 |
Source: https://artificialanalysis.ai/image/leaderboard/text-to-image/ People compare images made from the same prompt without knowing the model. Elo summarizes relative preference in this test; it is not a count of users or images. Model settings appear as separate entries, only covered models and prompts are represented, and overlapping rank ranges mean close places should not be treated as certain. Image editing is a different task.
The video track: The view changes with the task Artificial Analysis AA-Video-T2V v2.1, overall text-to-video with audio; accessed 3 November 2026:
| Rank | Model | Elo | Estimated rank range |
|---|---|---|---|
| 1 | Wan 3.0 | 1174 | 1–3 |
| 2 | Utopai X | 1168 | 1–5 |
| 3 | MiniMax H3 (768p) | 1162 | 1–6 |
| 4 | FLUX 3 Max | 1157 | 2–8 |
| 5 | Dreamina Seedance 2.5 | 1156 | 2–7 |
| 6 | HeyGen Video 1 | 1156 | 2–7 |
Evidence and responsible use
Source: https://artificialanalysis.ai/video/leaderboard/text-to-video Validated private evaluators compare videos generated from the same prompt without seeing the model names. The ranges overlap substantially, so the close scores do not establish a definitive champion. Silent video and image-to-video are separate tests and may produce different results. Method: https://artificialanalysis.ai/methodology/video
We found no comparable global model-use counts covering all platforms and APIs. Neither creative leaderboard should be described as the world’s five most used models. Each shows preferences within a particular test.
Make the comparison useful
Choose one task, such as a product shot or consistent character. Give the candidates the same prompt and reference where possible; note differences in duration, resolution and settings. Review several outputs without model names, then record waiting time, failed attempts, usage rights and what your own plan actually costs. API pricing is not the same as a subscription. Repeat with several scenes before deciding. This is a suggested test, not a test performed for this article.
Milo’s take: A public leaderboard can help you find candidates. Your finish line is the work you need to deliver; speed, continuity and cost may point to different tools.
