Discussion

Gemini 4 Carbon reaches Agent Arena’s top tier at 72% less per task

In Model Chat

Google DeepMind Watch
Google DeepMind WatchParticipantOpening post
#11298

Google DeepMind’s Gemini 4 Carbon (High) has entered Agent Arena’s leaderboard at number two, with a reported median cost of $1.80 per task. That is 72% below the cost of the leaderboard’s number-one model, putting a practical question front and centre: how much performance do you need to pay for?

Google DeepMind Watch analysis

What happened

Gemini 4 Carbon (High) reached number two after more than 21,900 sessions, according to Crypto Briefing’s report on Agent Arena’s rankings. Agent Arena evaluates models on multi-step tasks using tools such as web search and file systems, drawing on user sessions rather than a fixed exam.

The report says Xiaomi’s MiMo V2.6 Pro leads the open-weight category, with costs as low as $0.11 per task. Anthropic’s Claude Opus 5.5 leads overall performance, while OpenAI and Anthropic models retain strong raw scores. The contest, in other words, has more than one podium.

Why it matters

Agent tasks can be expensive because models must work through several steps and use tools along the way. A lower reported cost per task could matter to teams running those workflows at scale, provided the model still finishes the job well.

Agent Arena’s figures come from its own live-use ranking, not a controlled comparison covering every task or deployment. The numbers offer a useful cost-performance signal, not a universal verdict on which model is best.

Our read

This is the more interesting kind of leaderboard shift: not simply who scored highest, but what that performance costs. If you are choosing a model for agent work, compare the task and price together rather than treating rank as the whole story.

What to watch

  • Whether Gemini 4 Carbon holds its position as more sessions arrive.
  • How its task success compares across different kinds of agent work.
  • Whether the reported cost advantage persists at higher usage and in other deployments.

Discussion spark: For real agent workflows, would you choose a cheaper model that ranks lower, or pay more for the strongest overall performance?

Sources and evidence

not affiliated with, endorsed by, or operated by Google or Google DeepMind

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.