--- collection: tools slug: arena-ai locale: en title: "arena.ai, try and compare top AI models for free" summary: "A site that shows answers from two anonymous models to your prompt. You pick the better one, then see the names. We test it on writing, coding, images, video and Agent mode." publishedAt: 2026-11-05 status: live tags: ["tools","AI","Models","Comparison"] canonical: https://codebelkalam.com/en/tools/arena-ai --- # arena.ai, try and compare top AI models for free A site that shows answers from two anonymous models to your prompt. You pick the better one, then see the names. We test it on writing, coding, images, video and Agent mode. ## The short version arena.ai lets you type a prompt once and get replies from two models without knowing which is which. You pick the better result, and only then are the names revealed. The features below need you to sign in, and limits and availability change over time, so check the site itself. > **Warning:** > > The site is free because your conversations are shared with the model providers and may end up in public research datasets, according to the site's own FAQ. Do not paste confidential work or client data into it. ## How does it work? 1. You write your prompt. 2. Two replies appear side by side, from two models whose names you cannot see. 3. You vote for the better one. 4. The site reveals which model produced each reply. This is Battle mode. Its value is that you judge the output itself, not the brand or the model's reputation. In our test we asked for an explanation of an API in three lines, using a restaurant example, in Egyptian Arabic. We voted for the Claude model's reply because it kept to three lines, while the other model went over the limit. That is one prompt on one day, not a general ranking of either model. ```prompt Explain what an API is in three lines, in Egyptian Arabic, with a restaurant example. ``` ## What can you try? ### Writing and coding For coding, both models build a complete website in code, and you can try each one live on the page before you vote. You can click and scroll inside each site as you would on a real one, which shows differences that reading the code alone would not. ### Images and video The site generates images and videos from two different models so you can compare them. **Images.** We attached a photo and used the prompt below. We voted for `cobalt-grove-2` because the hair looked more realistic. That is our judgment on one image, not a general verdict on the model. ```prompt Create an extreme close-up underwater portrait after the subject has just jumped into a swimming pool or clear shallow water. Keep the mood calm, weightless, and ethereal, emphasizing light reflections. No struggle or urgency. ``` **Video.** We used the prompt below. We voted for `dreamina-seedance-2.0-720p` because it committed to the 9:16 requirement, while the other model ignored it. Judging only on format compliance is a narrow test: the quality of the clip is a separate question. The prompt is copied exactly as we typed it, including the Arabic on-screen text. ```prompt 8-second vertical 9:16 video for TikTok, native handheld screen-capture aesthetic (not glossy commercial). Scene: Dark matte slate digital canvas with a glowing statistical bell curve drawing itself rapidly. Subject & action: The dead center peak of the bell curve pulses dull gray and red, while the far left and far right tail edges ignite in bright metallic gold particles. Camera: Fast handheld push-in toward the glowing gold right edge of the curve. Lighting/mood: High-contrast dark-mode UI lighting, crisp and immediate. On-screen text: "إوعى تقبل أول إجابة من الـ AI!" Audio: Low digital hum and sharp UIlock sound ``` For video, the official help page says it requires a login, works in Battle mode and, at the time of writing, is limited to two generations a day. ### Agent mode In Agent mode you can ask for a presentation, a spreadsheet or a document and have it saved to your Google Drive. This needs the Drive, Docs and Slides connectors turned on for your Google account. In our test we asked for a five-slide Arabic deck. The agent created it and saved it to Drive, but it took about 22 minutes and several edit requests failed along the way. Do not assume it will be quick or smooth every time, and review the file before you use it. ```prompt Create a 5-slide Google Slides presentation in Arabic about: how to build your first website by talking to AI, and save it to my Drive. ``` > **Warning:** > > Before granting any connector access to Drive, read the actual consent screen and understand what you are allowing. Use an account without sensitive files while you experiment. ## The real ranking From user votes the site builds a leaderboard showing the strongest model for each kind of task. When we recorded the episode, the counter showed hundreds of millions of sessions. The leaderboard changes constantly, so use it to see where things stand now rather than relying on what is written here. The ranking reflects voters' preferences in blind matchups. It is a useful signal, not a final verdict on what suits your task. Test your own prompts. ## Try it yourself ### 1. Open arena.ai and sign in if you want to try video or Agent mode. Start with Battle mode and a simple text prompt. ### 2. Use a prompt from your real work after removing anything confidential or personal, and compare the two replies before you see the names. ### 3. Vote on a specific question: which followed my instructions? Which is more accurate? Then see who won. ### 4. Repeat with several different prompts. One result is not enough to judge a model. ## A note on accuracy Model names, usage limits and free features change quickly. This article reflects what we saw when recording the episode; the site itself is the source for current limits. ## Related lessons - [How do you choose an AI model?](/en/learn/choosing-an-ai-model): criteria for picking between models. - [How do you ask for exactly what you want?](/en/learn/what-is-a-prompt): describing the task and the result you want. - [Why does AI lie so confidently?](/en/learn/what-is-hallucination): why results need checking.