Most popular AI models: who leads on users, spend, tokens and downloads, and why

Part of our AI models guide: 11 posts on choosing and running AI models.
Ask which AI model is the most popular and you can get five right answers. ChatGPT has the most people: OpenAI says 1.2 billion use it every week. Google says its Gemini app has passed 1 billion users a month. Weekly and monthly counts measure different things, so the two do not compare directly. Businesses spend the most on Anthropic's Claude. Developers send the most text, counted in tokens, to cheap Chinese open models led by DeepSeek. The family people download and build on most is Alibaba's Qwen. And the models that top the quality tests are not the most used at all. It is like asking for the most popular car: the best seller, the busiest taxi and the one mechanics like to tune are rarely the same car.
Figures are as of 19 November 2026. A model is the AI itself; an app such as ChatGPT is how people reach it. A token is the smallest unit of text an AI model reads and writes, about three quarters of an English word. APIs, the web services that apps call, charge per million tokens: see cost per token explained. Open-weight means the maker publishes the model's files, so anyone can run it on their own GPUs (see open-weight vs closed models).
The leaders at a glance
Measure | Leader | Runner-up | Main reason it leads |
|---|---|---|---|
People using AI assistants | ChatGPT, 1.2 billion weekly | Gemini app, over 1 billion monthly | Habit, free tier, distribution |
Share of AI website visits, October 2026 | ChatGPT, 55.7% | Gemini, 24.2%; Claude, 11.5% and growing fastest of the three | First-mover habit |
US business API spend, 2026 | Anthropic, 55% | OpenAI, 31% | Strongest models and coding tools |
Developer tokens on OpenRouter, week to 18 November | DeepSeek V4.1 Flash (October update), 62.1 trillion tokens | Tencent's Hy4 flash, 46.6 trillion | Good enough at a fraction of the price, for automated tasks |
Downloads and fine-tunes, Hugging Face, last 30 days | Qwen, about 48% of downloads | Gemma, about 7% | Free for commercial use, many small sizes |
Quality leaderboards, mid-November | Claude Fable 5.2 (Artificial Analysis, 65); Gemini 4 Carbon (first on Arena) | GPT-6.2 Astra (Artificial Analysis, 62) | Highest test scores and people's votes |
Most used AI assistants
Assistant | Users | What is counted | Date | Source |
|---|---|---|---|---|
ChatGPT | 1.2 billion | Weekly users | 29 Sep 2026 | OpenAI, DevDay |
Gemini app | Over 1 billion | Monthly users | 28 Oct 2026 (first said 11 Aug) | Alphabet Q3 call |
Meta AI | Over 1 billion | Monthly, across Meta's apps | Oct 2025; no total since | Meta Q3 2025 call |
Doubao (ByteDance) | 419 million | Active users, China (period not stated) | Sep 2026 | QuestMobile |
Qwen app (Alibaba) | 158 million | Active users, China (period not stated) | Sep 2026 | QuestMobile |
DeepSeek app | 123 million | Active users, China (period not stated) | Sep 2026 | QuestMobile |
Grok (xAI) | About 117 million | Monthly users of Grok AI features | Mar 2026 | SpaceX IPO filing |
Claude | Not disclosed | Anthropic publishes no consumer total | n/a | See web share below |
Microsoft Copilot | Not disclosed for the consumer app | Last total: over 150 million monthly, for every Copilot, work tools included | Oct 2025 | Microsoft |
ChatGPT's weekly users rose from 800 million in October 2025 to 1.2 billion in September 2026, and Gemini's monthly users from about 650 million to over 1 billion in August 2026. On a monthly count, ChatGPT got there first: TNW reports it passed 1 billion monthly users on 3 June 2026, about ten weeks before Gemini.
Why they lead: ChatGPT has first-mover habit, the best-known name and a free tier. Gemini has distribution: it is built into Android phones, Search and Google's office apps. Meta AI sits inside apps people already open every day, and Doubao rides on ByteDance's Douyin, the Chinese version of TikTok.
Share of AI website visits
AI website | Oct 2026 | Sep 2026 | Nov 2025 |
|---|---|---|---|
ChatGPT | 55.7% | 56.4% | 69.5% |
Gemini | 24.2% | 25.0% | 15.9% |
Claude | 11.5% | 9.6% | 2.1% |
As of October 2026: Similarweb's share of visits to AI chatbot websites.
Similarweb counts websites only, so it misses use inside apps, such as Gemini on Android or Meta AI in WhatsApp. a16z's consumer AI app rankings (August data) also put ChatGPT first. (a16z invests in OpenAI.)
Where businesses spend
Measure | Anthropic | OpenAI | Source | |
|---|---|---|---|---|
Share of US enterprise LLM API spend, 2026 (2025) | 55% (40%) | 31% (27%) | 9% (21%) | Menlo Ventures, 3 Nov 2026 |
Share of US businesses on Ramp paying, October 2026 | 44.8% | 40.6% | Not listed | Ramp AI Index |
As of September 2026 (Menlo's survey of nearly 400 US enterprise decision-makers) and October 2026 (Ramp).
On Ramp, 56.8% of US businesses pay for some AI; xAI is at 6.3%, DeepSeek and Mistral at 0.2% each. Anthropic first passed OpenAI on this index in April 2026 (34.4% against 32.3%). Open-weight models (Menlo calls them open-source) carry 19% of enterprise tokens but take only 4% of API spend in Menlo's report, and Chinese models take nearly 90% of that spend: Kimi 61%, DeepSeek 15%, GLM 11% and Qwen 3% (rounded).
Why Anthropic leads: buyers told Menlo they rank capability above cost, and Menlo credits Claude's Opus models, which are strong at long tasks and coding (more in US AI models compared). Claude Code and OpenAI's Codex take 70% of enterprise coding spend. Buying also runs through channels companies already pay for. Microsoft turns Claude on by default for most Microsoft 365 commercial customers outside the EU, the UK, Norway, Switzerland, Iceland and Liechtenstein, under Microsoft's own data protection terms. Claude is also sold through Amazon Bedrock, Google's Gemini Enterprise Agent Platform (formerly Vertex AI) and Microsoft Foundry.
Read both with care: Menlo invests in Anthropic and OpenRouter and surveyed US firms only, and Ramp sees only its own customers.
Most used LLMs on OpenRouter
Rank | Model | Maker (country) | Tokens in 7 days, trillions |
|---|---|---|---|
1 | DeepSeek V4.1 Flash 1017 | DeepSeek (China) | 62.1 |
2 | Hy4 flash (new) | Tencent (China) | 46.6 |
3 | DeepSeek V4.1 Flash | DeepSeek (China) | 16.0 |
4 | Hy4 | Tencent (China) | 13.0 |
5 | GLM 5.3 Flash | Z.ai (China) | 10.9 |
6 | GPT-6 Luna | OpenAI (US) | 8.62 |
7 | Claude Opus 5.5 | Anthropic (US) | 7.63 |
8 | Step 5 Preview | StepFun (China) | 6.39 |
9 | MiMo-V2.6-Flash | Xiaomi (China) | 4.05 |
10 | MiniMax M3.1 Flash | MiniMax (China) | 3.67 |
As of the week to 18 November 2026.
OpenRouter is one API that reaches hundreds of models and ranks them by its own traffic. Eight of the ten come from Chinese labs (see Chinese AI models explained), and 1017 marks DeepSeek's 17 October update. Over the 30 days to 18 November, the two DeepSeek V4.1 Flash versions carried about 293 trillion tokens, about half of the top ten's 562 trillion. Chinese models carried about 88% of those top-ten tokens (our sums). Hy4 flash is new, and fresh releases can jump up the weekly list in their first days.
OpenRouter's figures are its own, and its users lean towards developers and price-sensitive apps.
Tokens are not money
In October, DeepSeek V4.1 Flash carried the most tokens in OpenRouter's coding tasks (23% for writing code, 24% for running commands). Claude Opus 5.5 led spend on most coding tasks; on OpenRouter's rankings it takes 24.5% of spend on implementation, against 8.5% for the runner-up, Kimi K3.1 Pro. The same gap runs across all paid traffic: in October, open-weight models carried most of OpenRouter's paid tokens but only about a quarter of its paid spend.
Why cheap models win tokens
Agents, software that calls a model over and over to finish a task, sent 76% of OpenRouter's 790 trillion tokens in October, up from 8% a year earlier. Their prompts repeat, and their median cache hit rate (the share of a prompt reused from earlier calls, usually billed at a discount) passed 90%. At that volume, price decides: OpenRouter's average price paid per million tokens in October was $0.084 for open-weight models against $0.52 for closed ones, about a sixth.
Model, top setting listed | Weights | Intelligence Index v5 | Cost per task |
|---|---|---|---|
Claude Fable 5.2 | Closed | 65 | $26.54 |
GPT-6.2 Astra | Closed | 62 | $4.00 |
GPT-6.2 Sol | Closed | 57 | $1.01 |
Kimi K3.1 Pro | Open | 46 | $3.68 |
DeepSeek V4.1 Flash 1017 | Open | 33 | $0.60 |
GPT-6 Luna | Closed | 28 | $0.15 |
From Artificial Analysis in mid-November 2026. The index combines many independent tests; cost per task is its average cost to run one test task on that model.
DeepSeek V4.1 Flash scores about half the leader's 65 for 44 times less per task ($0.60 against $26.54). In an agent loop that runs thousands of times a day, that is the trade developers make.
Most downloaded open models
Hugging Face is the main site for sharing open models. A fine-tune is a copy of a model retrained on extra data for one job, like an off-the-peg suit altered to fit. Size is counted in parameters, the numbers a model learned in training: more usually means smarter but needs more GPU memory. Apache 2.0 and MIT are licences that let anyone use a model commercially for free.
Family (maker) | Share of 30-day downloads | Share of new fine-tunes | Why |
|---|---|---|---|
Qwen (Alibaba) | About 48% | About 37% | Apache 2.0 on most sizes, from 0.6 to 27 billion parameters |
Gemma (Google) | About 7% | About 5% | Small sizes; Apache 2.0 since Gemma 4 |
DeepSeek | About 5% | About 2% | MIT licence, but mostly very large models |
Llama (Meta) | About 3% | About 5% | No new Llama since Llama 4 (April 2025); Meta's newer models carry the Muse name |
Nemotron (NVIDIA) | About 2% | About 1% | Open licence, several sizes |
gpt-oss (OpenAI) | About 1.5% | Under 1% | Apache 2.0, two sizes |
FlexiCloud's count from the Hugging Face API in mid-November 2026, rounded.
How we counted: downloads cover the top 1,000 text models and top 1,000 image-and-text models over the last 30 days, 710 million downloads in all. Families are matched by name, including other users' compressed (quantised) copies. GPT-2 (about 5%, mostly test traffic) is left out. Fine-tunes are the 25,798 of 104,159 new model uploads from 20 October to 19 November marked as retrained from a base model, and shares are of all 25,798.
Small models win downloads. In Hugging Face's Summer 2026 report, models under 1 billion parameters took about 83% of all-time downloads among models that state their size. Models above 70 billion took only about 3% of 2026 downloads. The same report counted 151,448 models built on Qwen, 2.6 times as many as Meta's.
Best is not most used
As the tables above show, Claude Fable 5.2 tops Artificial Analysis's Intelligence Index v5. On Arena's text leaderboard, where people vote between answers, Gemini 4 Carbon was first on 18 November with a score of 1526. Neither Fable nor Carbon is in OpenRouter's weekly top ten by tokens: most traffic goes to models that are good enough and far cheaper.
Models churn fast
About 1.3 new models reach OpenRouter every day, and coding work is much more likely than other tasks to run on a model released in the last 60 days. Loyalty is weakening too: OpenRouter says retention at the lab level fell almost across the board from 2025 to 2026. Still, of accounts with at least 100 paid requests, 78% stayed with Google from September to October and 74% with Anthropic and with OpenAI.
How popular models reach businesses
Open models are often served by companies other than the labs that made them. Inference providers rent or buy GPU servers, load open-weight models such as DeepSeek V4.1 Flash, and sell access through an API, priced per million tokens. Closed models such as Claude, GPT and Gemini are sold by their makers and through the big clouds' AI platforms.
For a business, that means no GPUs to buy, no servers to keep running and no model upgrades to manage. You pay only for the tokens you use, and repeated prompts that hit the cache cost less. You can often switch models by changing one line of code, which matters when the leaders change every month.
The sums can flip in two cases. At steady, high volume, a server rented for a year can cost less than the same tokens bought one call at a time. And some data must stay on hardware you choose. Our guides cover how inference providers sell tokens and whether to rent, buy or use an API.
Choosing a model: pick your race first
- Match the measure to your workload. A consumer chat app, an office assistant and a high-volume agent are different races, with different winners.
- For agents and coding, compare the cost of a finished task, not the price per token.
- Test on your own tasks every month, and build so you can swap models.
- If an open-weight leader such as DeepSeek V4.1 Flash, GLM-5.3 or Qwen3.8 passes your quality bar at steady volume, running it on your own servers can cost less than paying per token, but count the engineers who keep it running. Size it with our GPU sizing guide.
When an open model earns its own server
Most teams should start on an API. When one open-weight model carries steady volume, a dedicated server can cost less per token, and that is where we come in: we source whole 8-GPU servers, H100 to B300 (GB300 NVL72 racks on request), from verified operators in North America, the UK, Europe, India, Asia Pacific and the Gulf, on terms of a year or more paid upfront. Send us the model and your monthly token volume and we will work out which server fits, or see the GPU servers we source.
Where these figures come from
- ChatGPT: OpenAI DevDay recap (29 September 2026), as reported by TechCrunch; 800 million weekly from OpenAI's DevDay 2025 (6 October 2025); the monthly comparison: TNW (12 August 2026).
- Gemini: Alphabet Q3 2026 remarks (28 October 2026); about 650 million monthly from Alphabet's Q3 2025 call (29 October 2025).
- Meta AI: Meta Q3 2025 remarks (29 October 2025). Copilot: Microsoft FY26 Q1 call (29 October 2025). Grok: TechCrunch on the SpaceX filing (20 May 2026).
- China apps: QuestMobile, via TMTPost (21 October 2026).
- Web share: Similarweb, via OfficeChai (10 November 2026). Consumer apps: a16z Top 100, 7th edition (5 October 2026).
- Business spend: Menlo Ventures, State of Enterprise AI (3 November 2026); Ramp AI Index (data through October 2026) and its May 2026 report (April data). Claude in Microsoft 365: Microsoft Learn (4 November 2026).
- OpenRouter: rankings (data through 18 November 2026, including the spend shares by task; 30-day shares are our sums), October 2026 data and Model Half-Life (1 October 2026).
- Quality: Artificial Analysis leaderboard (19 November 2026) and Arena text leaderboard (18 November 2026).
- Downloads: our count from the Hugging Face API (19 November 2026) and Hugging Face's State of open models, Summer 2026 (14 August 2026).