Gemini 4 Carbon Tops Artificial Analysis Intelligence Index After Google DeepMind Launch

Google DeepMind released Gemini 4 Carbon on October 28, 2026, positioning the model as its strongest frontier system yet. Announced in an official blog post, the release came two weeks after Gemini 4 Argon reached general availability. The new model leads the updated Artificial Analysis Intelligence Index v5.0 with a score of 59, ahead of Claude Opus 5.5 and GPT-6 Astra.

Gemini 4 Carbon Performance on Artificial Analysis Intelligence Index and Cybersecurity Benchmarks

Carbon achieved the highest ranking on the Artificial Analysis Intelligence Index v5.0 by scoring 59 in the high-reasoning tier. It placed first in four of nine evaluations, including HLE-Diamond and Terminal-Bench-Science. The model also recorded an 89-point Elo gain on WebDev Arena, reaching approximately 1768 and fourth place overall.

Google reported cybersecurity results measured without guardrails. Carbon scored 88.2 percent on SEC-bench Pro for vulnerability detection, 54.8 percent on ExploitGym for exploitation tasks, and 71.0 percent on CWE-bench for code patching. Internal tests showed more than double the secure code generation rate compared with Argon and zero successful indirect prompt injection attacks on Gray Swan benchmarks.

Additional gains appear in agentic coding, long-horizon planning, scientific reasoning, and multimodal understanding. The model reduced instances of misalignment and fabrication while handling ambiguous prompts more reliably. Pricing matches the introductory rate set for Argon at two dollars per million input tokens and ten dollars per million output tokens, with cached input offered at a 95 percent discount.

Staged Rollout, Context Window, and Comparison With Gemini 4 Argon

Access begins with a limited group of cybersecurity defenders through the Fairwind Program during the week of announcement. Google plans gradual expansion to additional trusted partners before wider availability to developers and enterprises. The company is also participating in voluntary pre-release testing coordinated by U.S. government agencies.

The context window remains one million tokens, consistent with Argon. Native multimodal support covers text, image, video, and speech inputs with text outputs. No parameter count or detailed architecture information was disclosed. Pricing will rise to four dollars per million input tokens and twenty dollars per million output tokens after the introductory period ends.

Carbon builds directly on Argon with targeted improvements drawn from internal testing and early user feedback. The two-week interval between releases reflects an accelerated iteration cycle focused on agentic workflows and cybersecurity applications. Argon itself moved from announcement to broader availability in roughly fourteen days, providing a reference timeline for Carbon’s expected expansion.

Developer and Enterprise Implications

Organizations evaluating frontier models for secure code generation and long-running agent tasks now have updated benchmark data to compare against Claude Opus 5.5 and GPT-6 Astra. The staged rollout means most developers will continue using Argon until Fairwind feedback informs guardrail adjustments. Pricing stability during the introductory window allows teams to run extended evaluations before the rate increase takes effect.

Google has not confirmed an exact date for wider public or developer access. Teams planning integration should monitor Fairwind participant updates and subsequent announcements on guardrail refinements. The combination of top index ranking and documented cybersecurity gains positions Carbon as a candidate for workloads that require both high reasoning performance and reduced attack surface.