Skip to content
Advertisement
Google DeepMind’s Gemini 4 Carbon takes the top spot in LM Arena’s Text Arena

Google DeepMind’s Gemini 4 Carbon takes the top spot in LM Arena’s Text Arena

Photo: Tim Witzdam / Pexels

The new Gemini variant scored 1,528 points on the blind-voting leaderboard and placed in the top four on agent and web development tests

Google DeepMind has a new leaderboard champion. Gemini 4 Carbon, a variant in the Gemini 4 generation, now holds the number one spot in LM Arena’s Text Arena with a score of 1,528 points.

Carbon’s showing goes beyond text. The model also ranked second in Agent Arena and fourth in WebDev Arena, so this isn’t a one-trick result.

How Gemini 4 Carbon stacked up

The headline number is 1,528 points in Text Arena, enough for the top global ranking. Text Arena measures how people rate a model’s written answers when compared directly against rival models.

Second place in Agent Arena may matter even more over time. That category tests multi-step and autonomous work, where a model has to plan, act and follow through without constant hand-holding.

The fourth-place finish in WebDev Arena rounds out the picture. That leaderboard rewards models that can produce working web code and functional applications, not just plausible-looking snippets.

Advertisement

The “Carbon” label marks a specific variant within the larger Gemini 4 family.

Why LM Arena scores carry weight

Static benchmarks use fixed sets of questions. Once those questions circulate widely, models can end up effectively studying for the test, which inflates scores without improving real-world usefulness.

LM Arena takes a different approach. It shows users two anonymous responses to the same prompt and asks which one is better, then feeds those votes into an Elo-style rating system.

Put simply, a model’s score rises when people prefer it over opponents in blind matchups, and falls when they don’t. Nobody knows which brand they’re voting for until after the fact.

The platform draws on millions of these human head-to-head votes. That scale, combined with the blind format, is why the leaderboard has become a widely cited reference point for frontier model performance.

The Gemini family’s fast climb

Carbon arrives after a rapid run of upgrades across the Gemini lineup, which has moved from Gemini 1.0 through successive generations to Gemini 4.

The competition includes OpenAI, Anthropic, Meta and xAI, all of which ship new models on a relentless schedule.

What this means for the AI race

For Google, the result is a useful counter to the narrative that it spent years playing catch-up in generative AI. A first-place Text Arena finish gives the company a concrete, third-party-style data point to point to.

There are caveats worth keeping in mind. Human preference votes capture what people like reading, which doesn’t always line up perfectly with accuracy, safety or cost efficiency at scale.

The WebDev result is a reminder of that nuance. Carbon leads on text, but other models still edge it out on building web applications, so the “best model” question depends heavily on the job.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.