FAQ
Frequently asked questions
Today’s answers
What is the most-used AI model right now?
As of Nov 22, 2026, Hy4 flash is the most-used AI model on OpenRouter, with 103T tokens in the last 7 days. Next are DeepSeek V4.1 Flash 1017 and DeepSeek V4.1 Flash.
What is the best-performing AI model right now?
On the Arena text leaderboard released Nov 19, 2026, Gemini 4 Carbon ranks #1 with a score of 1527 from 4,262 blind votes. Next are Gemini 4 Argon and Gemini 4 Neon.
Which AI model is the best value?
Among the top-performing models, MiMo-V2.6-Pro is the cheapest ($0.87 per 1M tokens), with an Arena score of 1481.
What is the most-used AI model for coding?
In the week of Nov 9 – 15, DeepSeek V4.1 Flash 1017 was used most for programming on OpenRouter, followed by DeepSeek V4.1 Flash.
What is the best AI image model?
On the Arena image leaderboard released Nov 20, 2026, GPT Image 2.5 Sunburst ranks #1 for quality with a score of 1429.
What is the best AI video model?
On the Arena video leaderboard released Nov 9, 2026, Gemini Omni 1.2 Flash ranks #1 for quality with a score of 1540.
What is the most popular open-weight model right now?
Among official releases on Hugging Face, Qwen3.8 27B by Alibaba is the most popular open-weight LLM as of Nov 22, 2026. It needs about 17 GB of memory at 4-bit.
How the rankings work
What is Avenvote?
A free site that ranks AI models every day using public data: the most-used, the best-performing and the best-value language models, plus image and video models, in 9 languages.
How often are the rankings updated?
Every day around 00:10 UTC. Usage covers the last 7 days; performance and quality change whenever Arena publishes a new leaderboard (every few days to weeks).
Where does the data come from?
Usage comes from OpenRouter, performance and quality from the Arena leaderboard dataset, and image and video usage share from Vercel AI Gateway (all CC BY 4.0). Prices come from the OpenRouter model list.
What does “usage” mean?
The number of tokens developers and apps sent and received through OpenRouter in the last 7 days. It measures API use, not the number of people using apps like ChatGPT or Gemini.
How is performance measured?
With Arena blind votes: people compare two anonymous answers to the same prompt and pick the better one. We use the style-controlled score, so formatting doesn’t inflate results, and show one entry per model among the top models available on OpenRouter.
How is “best value” calculated?
We take the models in the performance ranking and sort them by output price per 1M tokens, cheapest first, so strong models that cost less stand out.
What is a token?
The unit AI models use to read and write text. In English, one word is about 1.3 tokens. Prices are per 1 million tokens, with separate prices for input (what the AI reads) and output (what it writes).
Why are prices shown in my currency?
We convert OpenRouter’s USD list prices at fixed exchange rates. What you actually pay depends on the provider, caching and discounts.
Why isn’t a model listed?
Usage shows the top 30 on OpenRouter (top 30 per use case), and performance shows the top Arena models that are also on OpenRouter. Models available only in their own apps may not appear.
What does “Undisclosed maker” mean?
Labs sometimes test new models on OpenRouter under a code name before launch. We show “Undisclosed maker” until the maker is announced.
Are the official-site links ads?
They link to each model’s official site. If a link is ever an affiliate link, it is labeled, and it never affects the rankings.
How do recommendations work?
Anyone can recommend each model once a day, with no account. We don’t store IP addresses; a one-way code that only stops repeat votes is kept for 8 days.
Can I use or cite the data?
Yes, for quotes, screenshots, posts, articles and AI answers — please link to the page you used. Copying our compiled data in bulk or redistributing it as a dataset needs our permission. The source data (OpenRouter, Arena, Vercel) stays under CC BY 4.0, and you can get it from those sources directly.
How do you estimate the memory an open-weight model needs?
From the parameter count on Hugging Face: about 0.55 GB per billion parameters for LLMs at 4-bit and 1.1 GB for image and video models at 8-bit, plus 10%. A Mac can use about 70% of its unified memory and a PC without a GPU about 60% of its RAM. Real needs change with context length and settings, so treat it as a rough guide.
How do I report a data error?
Email [email protected] or use the Contact page.