Google DeepMind officially announced Gemini 4 Carbon on October 28, 2026, marking its second frontier model release in a single month. The proprietary model immediately claimed the top spot on the independently run Artificial Analysis Intelligence Index v5.0 with a score of 59, edging out Anthropic’s Claude Opus 5.5 by a single point. However, the headline benchmark victory is less significant than the distribution strategy, which prioritizes cybersecurity defenders over general users.

Composite Wins Hide Task-Level Losses

While Carbon posts a nine-point jump over its predecessor, Gemini 4 Argon, the composite lead of one point over Claude Opus 5.5 (max) and two points over GPT-6 Astra (max) masks critical performance gaps. On the AA Coding Agent Index, Carbon scores 59 compared to Opus 5.5’s 62, indicating that best-in-class aggregate metrics do not guarantee superiority in specific engineering tasks. It also trails Kimi K3.1 Pro, Muse Wonder, and Grok 4.7 on the AA-Briefcase benchmark, reinforcing that model selection must remain task-specific.

Security Benchmarks and Guardrail Implications

The launch is heavily framed around cybersecurity, with Google reporting strong performance on vulnerability detection and patching. Carbon achieved 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench. Crucially, these security benchmarks were run without guardrails, a choice that highlights why access is restricted. Google also noted that cheating on Terminal-Bench 4.0 collapsed to 0.0% successful and 0.4% attempted, suggesting improved evaluation integrity alongside rising accuracy.

The Defender-First Rollout Strategy

Pricing has been announced at $2/$10 per million tokens for introductory use, rising to $4/$20 later, with cached input discounted by 95%. However, the API is not yet live for the general public. Instead, a subset of cyber defenders receives access this week, followed by vetted partners through the Fairwind Program. Ars Technica summarized the current state bluntly: you can't use. This staged approach, including government pre-release review, establishes a new template for frontier capability distribution.

Key Takeaways

  • Gemini 4 Carbon leads the Artificial Analysis Index with 59 points but trails Opus 5.5 in coding agents.
  • Security benchmarks (88.2% on SEC-bench Pro) were run without guardrails, justifying restricted access.
  • The rollout is staged: defenders first, then Fairwind partners, with no public API date.
  • Current cost per task is $7.76, higher than GPT-6 Astra ($6.25) but lower than Opus 5.5 ($10.04).

The Bottom Line

Builders should stick with Gemini 4 Argon for now, as the staged, defender-first access to Carbon is a strategic signal that frontier AI distribution is becoming more controlled and less immediately available.