Google Gemini 4 Carbon Claims Top AI Benchmark Spot Two Weeks After Argon

Quick Facts

• Gemini 4 Carbon scored 59 on the Artificial Analysis Intelligence Index v5.0, ranking first ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57.
• The model hit 88.2% on SEC-bench Pro and 54.8% on ExploitGym, both state-of-the-art results for cybersecurity evaluations.
• Google released Carbon just two weeks after the wider rollout of Gemini 4 Argon, signaling a faster cadence of frontier model releases.

Google DeepMind announced Gemini 4 Carbon on October 28, 2026, calling it its strongest model yet for cybersecurity and long-horizon tasks. The release came 14 days after Gemini 4 Argon reached general availability.

Alphabet CEO Sundar Pichai confirmed the pace on X: "I did say we'd be releasing our models periodically. Carbon shows the pace at which we're pushing the frontier." The official announcement was authored by Koray Kavukcuoglu, SVP at Google DeepMind and the company's Chief AI Architect.

Benchmark Performance

Independent evaluator Artificial Analysis placed Carbon first on its Intelligence Index v5.0, a rescaled index built from nine evaluations that launched the day before Carbon's announcement. The new scoring methodology makes direct comparisons to previous index versions unreliable.

Carbon leads four of the nine index categories: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA. It also tops 15 of the 17 rows in Google's own benchmark table compared to Argon and competing models.

Two gaps remain. GPT-6 Astra holds the lead on FrontierSWE 2 software engineering at 65.5%, against Carbon's 61.2%. Claude Opus 5.5 from Anthropic leads on OSWorld-2.1 computer-use at 87.2%, against Carbon's 81.5%.

On coding agents, Carbon scores 59 on the Artificial Analysis Coding Agent Index using the Antigravity CLI. That puts it second, three points behind Claude Code running Claude Opus 5.5 at 62. Carbon ranks second on Terminal-Bench 4.0 at 64% and on FrontierCode v1.1 at 51%.

Cybersecurity Is the Lead Story

Google built Carbon with a clear focus on security work. The model scored 88.2% on SEC-bench Pro, which tests autonomous discovery of complex flaws in real-world software, and 54.8% on ExploitGym, an exploitation benchmark. It also posted 71.0% on CWE-bench for code patching.

Internal testing showed Carbon generates secure code at more than twice the rate of Argon. High-severity misalignment alerts in live coding environments with internal teams fell by nearly 80% compared to the previous model.

Google's Chrome team provided an additional data point. Kavukcuoglu wrote that the team reported Carbon "performed more than 10X better than another lab's large frontier model when generating high-quality patches."

Google Cloud CEO Thomas Kurian pointed to rising enterprise demand as context, saying cybersecurity risks are driving customers toward AI-powered security platforms.

Safety and Reliability Gains

Google reported measurable improvements in model reliability alongside the security numbers. Carbon reduced successful cheating attempts on Terminal-Bench 4.0 to 0.0%, with overall cheat attempts falling to 0.4%. On Gray Swan's indirect prompt-injection benchmark, Carbon scored 0.0% attack success rate. Argon had scored 0.7% on the same test.

Kavukcuoglu also noted that Carbon cuts unsupported claims by approximately 2x compared to Argon, a signal that the model hallucinates less in text generation tasks.

Web Development and Access

On WebDev Arena, Carbon gained 89 Elo points over Argon to land at 1768, placing fourth on that leaderboard.

Access to Carbon is currently limited. Google has not announced broad general availability, and the model is being offered as a preview. Pichai described the release as part of a planned cadence, suggesting additional models will follow at a similar pace.