Many MindedGemini 4 Carbon tops an AI index two weeks after its predecessor
Google DeepMind has released Gemini 4 Carbon, about two weeks after Gemini 4 Argon, according to an October 28, 2026 report. Carbon scores 59 on the Artificial Analysis Intelligence Index v5.0, up from Argon's 50, and ranks first. It leads four of the index's nine individual evaluations at its high reasoning tier. Pricing is the same as Argon.
Google positions Carbon for long-horizon engineering and multi-day autonomous coding, and says internal teams ran it for several days straight. Results elsewhere are mixed. In agentic coding on Antigravity it rose from 53 to 59 but trails Claude Code with Opus 5.5 at 62. On AA-Briefcase knowledge work it gained 126 Elo to 1,609 but sits behind Kimi K3.1 Pro, Muse Wonder and Grok 4.7. Answer accuracy there rose from 48% to 58%, while hallucination rose from 4% to 11%. On Arena.ai it leads Text Arena and places second in Agent Arena, with a median task cost of $1.80, 72% below Opus 5.5. Vals AI ranks it second at 70.36%, at $10.51 per test.
For abundance, the relevant signal is cost per unit of capable work. Stronger agentic performance at unchanged pricing, and a low median task cost in Agent Arena, point toward cheaper delegated knowledge work. But Carbon produces 2.3x the reasoning tokens and 64% more output tokens than Argon, so real cost per task depends on usage. Nothing here changes who can access it.
Caveats: Google reports first place in 15 of 16 benchmarks it chose, and its claims about secure code and Chrome patch writing cannot be independently verified. The prompt injection benchmark, where attack success fell from 0.7% to 0.0%, appears saturated and does not mean the problem is solved. Carbon is in U.S. government voluntary pre-release testing and will first go to selected cybersecurity experts. The source's access section was truncated.
Source: The Decoder