Gemini 4 Carbon debuts #1 in Text Arena and #2 in Agent Arena
From the AI digest for Oct 30, 2026
What happened: Google DeepMind's Gemini 4 Carbon made its first appearance on Arena's boards, and Demis Hassabis amplified the results (Arena.ai via Hassabis on X). It enters at #1 in Text Arena (1,528 points, ahead of Gemini 4 Argon at 1,520) and #2 in Agent Arena with a +16.2% net improvement across roughly 21,900 real-world agentic sessions from Arena's community — placed ahead of GPT-6 Astra and Claude Fable 5.1. Agent Arena reports a $1.80 median cost per task, 72% less than the board's #1, Claude Opus 5.5 (Max). In Code Arena: WebDev it scores #4 (1,768 points, two behind Claude Sonnet 5.5 xHigh, 89 ahead of Gemini 4 Argon). Standout signals: #1 in steerability (+18.6%) and #3 in confirmed user success (+18.7%). Hassabis restated the positioning: Carbon is "state-of-the-art across cybersecurity and long-horizon capabilities."
What's genuinely new: the first independent board placements for Carbon, which until now had only closed-benchmark claims (an Artificial Analysis v5.0 index score of 59) and gated Fairwind access for a subset of cyber defenders. Session-level agent results and a cost-per-task comparison against Opus 5.5 (Max) are new specifics. One caveat: the numbers arrive through Carbon's own team's amplification of Arena's post — no independent writeup of the board results in this fetch.
Watch: whether access widens beyond Fairwind cyber defenders; whether broader independent testing reproduces the #1 placements; and any API or pricing availability announcement.
Links: Demis Hassabis on X · Oct 28, 1:24 PM ET