Quick answer
To test unreleased AI models in Arena, go to arena.ai, open battle mode and send a prompt. Two anonymous models answer side by side, you vote for the better answer, and the model names are shown after you vote. Pre-release checkpoints, such as the new Gemini 4 Carbon, appear only occasionally, so run several rounds. Use your own real tasks as prompts, keep a simple log of results, and never paste private or client data.
Key takeaways
- Battle mode hides model names until you vote, which keeps your judgement fair
- Unreleased checkpoints turn up at random, so expect to run many rounds
- Test with tasks from your real work, not trick questions
- Keep a simple log of the prompt, both answers and the winner
- Never paste private, client or customer data into a public comparison site
- Treat a checkpoint as a preview: it can change before release
📘 Main article: This guide supports Claude Duet: Anthropic’s Proactive AI Agent, Plus Gemini 4 Carbon’s New Checkpoint. It shows you how to try the anonymous models mentioned in that article, such as the new Gemini 4 Carbon checkpoint, for yourself.
In this guide
New AI models often appear in public testing weeks before they launch. If you know where to look, you can try them early and decide for yourself whether the hype is justified.
This guide answers one question: how do you test unreleased AI models in Arena battle mode, and judge them fairly? It supports our main article, Claude Duet: Anthropic’s Proactive AI Agent, Plus Gemini 4 Carbon’s New Checkpoint, which is based on WorldofAI’s latest news video.
What is Arena battle mode?
Battle mode is a blind test: you send one prompt, two hidden models answer, and you pick the better one. Arena, at arena.ai, is a public site that ranks AI models using these votes.
AI labs sometimes put unfinished versions of their models, called checkpoints, into Arena under code names. That is how people spotted the new Gemini 4 Carbon checkpoint and a model called “Kibbitz” that claims to be Kimi.
Why test models before they are released?
Because benchmarks and demos rarely match your actual work. A model that builds a stunning 3D scene may still write clumsy sales emails.
- You see real strengths and weaknesses on tasks you care about
- You are ready to switch, or not, on launch day
- You avoid buying into hype based on cherry-picked clips
- It is free, so it costs only your time
How do you run a battle in Arena?
The steps are simple, and the main skill is patience.
- Open arena.ai and choose battle mode. If you want coding or web design tests, look for the coding section of the site.
- Paste your test prompt and send it.
- Read both answers in full before deciding. Check facts, tone and whether it followed every instruction.
- Vote for the better answer, or call it a tie.
- Note the model names that appear after your vote.
- Start a new round with the same prompt to meet different models.
💡 Tip: Use the same prompt across many rounds. That way, when a checkpoint you are curious about finally turns up, you can compare its answer with every other model you have already seen.
What prompts should you test with?
Use short versions of tasks you already do every week. Here are examples for a small online business:
| Task type | Example prompt idea | What to check |
|---|---|---|
| Rewrite this (made-up) product launch email in a friendly British tone, under 150 words | Tone, length, clear call to action | |
| Landing page | Build a single-page HTML sign-up section for a free checklist | Does it work on mobile, is the layout clean |
| Planning | Create a four-week content plan for a beginner affiliate blog | Realistic, specific, no invented statistics |
| Graphics | Draw a simple SVG icon of an envelope with a lightning bolt | Clean shapes, sensible colours |
| Reasoning | Compare two pricing options for an online course and recommend one | Clear logic, states its assumptions |
⚠️ Warning: Treat anything you type into Arena as public. Prompts on public comparison sites may be stored or shared for research. Use made-up names and figures, and never paste client details, customer lists, passwords or API keys.
How do you record and judge the results?
A simple spreadsheet is enough. Add one row per round with these columns:
- Date and prompt used
- Model A and model B names (after voting)
- Winner, and one sentence on why
- Any errors, such as made-up facts or ignored instructions
After ten or more rounds, patterns appear. You might find one model wins on writing and another on code. For a more structured method, see Cheap vs Premium AI Model: How to Run a Fair Side-by-Side Test.
Curious what the fuss is about? Read our main article, Claude Duet: Anthropic’s Proactive AI Agent, Plus Gemini 4 Carbon’s New Checkpoint, for the checkpoints worth looking out for.
How can you tell which model you got?
Arena reveals the names after you vote, but unreleased checkpoints often use code names. Linking a code name to a real product usually relies on community detective work, such as a model describing itself, or a lab confirming it later.
🔎 Unconfirmed / rumour: A model saying “I am Kimi” or “I am Gemini” is not proof. Models can be wrong about themselves, and code names change. Treat any identification as a guess until the lab announces it.
What mistakes should you avoid?
- Judging from one round. Every model has good and bad answers.
- Using trick questions only. They rarely reflect real work.
- Skimming long answers. Errors often hide in the middle.
- Assuming the checkpoint is the final product. Labs keep changing models before launch.
- Pasting real customer data. Keep tests made up.
Frequently asked questions
Is Arena free to use?
Yes, you can compare models in battle mode for free. Some features may need you to sign in.
Can I choose to test Gemini 4 Carbon directly?
Not while it is unreleased. Pre-release checkpoints appear at random in battle mode, so you may need many rounds to meet one.
Does my vote matter?
Yes. Arena uses votes from many people to rank models, so honest votes on real tasks help everyone.
Is it safe to paste business information?
No. Treat everything as public and use made-up examples instead of client or customer data.
How many rounds should I run?
Aim for at least ten with the same prompt set before drawing any conclusions.
The bottom line
Arena battle mode lets you try new and unreleased AI models for free, judged blind on your own tasks. Use real, made-up-data prompts, run plenty of rounds, log the winners and treat code-named checkpoints as previews. Then read our main article for the models worth watching this week.
Related posts
- Cheap vs Premium AI Model: How to Run a Fair Side-by-Side Test
- GPT-6.5 Leak? MiniMax M3.1 Flash, Gemini 4 Carbon and Kimi K3.1 Explained
- Gemini 4 Argon Review: Has Google Finally Caught Up?
[…] How to Test Unreleased AI Models in Arena Battle Mode — how to try anonymous and pre-release models yourself, and how to judge them fairly. […]
[…] To compare the money side properly, use How to Work Out the Real Cost of an AI Task Before You Switch Models. To try unreleased models for free, see How to Test Unreleased AI Models in Arena Battle Mode. […]