LLM Battle Arena

How scoring works

🎤 Efficiency: how hard the crowd cheers per fan. A small club that goes wild beats a stadium that's checking its phone. (0.67% ≈ about 7 reactions per 1,000 followers on an average post.)

🏟️ Reach: how loud the whole stadium got this week, counting every like, repost and reply over the last 7 days.

Small print: Reactions = likes + reposts + replies, so one fan can count three times. It's not a head count. Updated every few hours from public X posts.

Engagement rate

The Efficiency board ranks by engagement rate. A reaction is a like, a repost, or a reply. A quote is not added as an extra reaction. When a quote is an original post, not a reply and not a repost, it counts as its own post, the same as any other post.

For each account, add the reactions on its counting posts from the last 7 days, divide by the number of posts, then divide by followers. Shown as a percent, that is the average reactions per post divided by followers, times 100. A figure of 6.7 reactions per 1,000 followers is an engagement rate of 0.67%.

This counts reactions, not unique people. One person can like, repost, and reply, and each of those counts. It is an engagement rate. It is not the percent of followers who reacted.

Companies use the official lab and product accounts only. A company's rate is the average of those accounts, weighted by how many posts each one made, so a busier account counts more. People are not mixed into the company rate.

Leaders are scored on their own, with that same rate. Companies are ranked against companies. Leaders are ranked against leaders.

The bar and the arrows

The thin bar shows distance to #1. The leader fills the bar. An account at half the leader's rate fills half of it. The bar is only a picture. It has no number of its own.

An arrow is the change in the engagement rate, in percentage points, and the change in rank. ▲ +0.04 pts means the rate rose by 0.04 percentage points. ▲ 2 places means the account moved up two ranks. Rank arrows only include accounts that were ranked both times. Someone who drops below the post minimum does not count as anyone else climbing, and a new account does not count as anyone else falling. A zero change is hidden. If that account has no earlier number, the arrow is hidden too. Nothing is filled in.

When the saved history reaches back 7 days, the arrows compare with last week. Until then, they compare with the last update. The page says which one it is. The biggest-mover line is the largest engagement-rate change, in percentage points, among accounts ranked both times. If that change rounds to 0.00, the line stays hidden. A tie goes to the larger rank gain.

Reach

Reach is the plain total of those same reactions over the same 7 days. It is not divided by followers or by the number of posts. The biggest total sits at the top. A large account can lead Reach and still sit lower on Efficiency.

Reach arrows use the same saved history. They show the change in the reaction total and in rank. They stay hidden until a snapshot has stored that total.

The 7-day window

Both scores use a rolling 7 days. A post older than that drops out. Hot today is separate: it ranks total reactions from the last 24 hours, and it does not change either weekly score.

Which posts count

Official lab and product accounts count every post. That includes accounts such as @AnthropicAI, @OpenAI, @GoogleDeepMind, @SpaceXSI, @grok, @GeminiApp, and @OpenAIDevs.

Individual leaders, and general corporate accounts such as @Meta, only count posts that are about AI. The check looks for whole words, ignores capitals, and treats a mention such as @bot as a match. A post that does not match is left out of both scores. It also stays out of Top post, Hot today, and the post list on a card.

Anyone with fewer than 3 counting posts in the week is not ranked. A leader in that spot says "not enough AI posts this week."

Updates and source

The numbers refresh about every 3 hours. The posts are public posts on X.