SYLEN
AboutNewsConferenceMembershipDonate

Email updates

Conference, news, and membership updates by email.

Site

  • About
  • News
  • Membership
  • Waitlist
  • Donate

Conference

  • Conference 2027
  • Call for papers

Account

  • Create account
  • Membership details

SYLEN

  • Guidelines
  • Privacy
  • Terms

© 2026 Systems Leadership and Engineering Network. sylen.org.

Membership details →
Back to news
General SESource: artificialanalysis.aiOctober 29, 2026

Google Releases Gemini 4 Carbon, Securing Top Spot on Artificial Analysis Intelligence Index v5

Google’s Gemini 4 Carbon has claimed the top position on the Artificial Analysis Intelligence Index v5 with a score of 59, demonstrating significant performance gains in complex reasoning while lowering output token consumption compared to competing frontier models. Standardized at $4/$20 per million input/output tokens, Carbon ranks second on the Coding Agent Index behind Claude Opus 5.5 while maintaining an 11% hallucination rate on factual evaluation sets.

Benchmark Performance and System Capabilities

Google’s Gemini 4 Carbon has taken the #1 rank on the Artificial Analysis Intelligence Index v5, earning an overall intelligence score of 59. The model achieved top scores in four of the nine individual evaluations comprising Index v5: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA.

Carbon supports a 1M token context window with multimodal input processing across text, image, video, and speech, producing text output. Access is currently limited to selected users during an initial rollout. Google's standard pricing tier is set at $4.00 per 1M input tokens and $20.00 per 1M output tokens, with cached input tokens discounted by 95% to $0.20 per 1M tokens. A temporary 50% launch discount is active without a published end date.

Coding Agent Benchmark Results

Paired with the Antigravity CLI, Gemini 4 Carbon (high reasoning) scored 59 on the Artificial Analysis Coding Agent Index. It ranks second overall behind Claude Code running Claude Opus 5.5 (max), which scored 62.

  • Terminal-Bench 4.0: 64% execution accuracy.
  • FrontierCode v1.1: 51% accuracy, behind Claude Opus 5.5 (max) at 56%.

This trajectory represents a measurable shift for Google's model family, which had historically lagged behind competitor architectures in agentic coding tasks.

Agentic Workloads and Hallucination Dynamics

In agentic knowledge workflows benchmarked by AA-Briefcase, Gemini 4 Carbon (high) recorded an overall score of 1609 Elo—a 126-point gain over Gemini 4 Argon (high, 1483 Elo). Sub-category performance gains were led by Presentation, which rose 293 Elo to 1598, and Analytical Quality, which gained 120 Elo to 1671. Despite these increases, Carbon continues to trail several models on AA-Briefcase, including Grok 4.7 (xhigh, 1638 Elo), Muse Wonder (max, 1659 Elo), and Kimi K3.1 Pro (1678 Elo).

On the updated AA-Omniscience v1.1 dataset, Carbon reached a top score of 54, leading Claude Opus 5.5 (max, 48) and Gemini 4 Argon (46). Its task accuracy rose 10 percentage points over Argon to 58%. While Carbon's hallucination rate increased from Argon’s 4% up to 11%, it remains the second-lowest recorded rate on the index, significantly below the 50% hallucination rates observed in both Claude Opus 5.5 (max) and GPT-6 Astra (max).

Token Efficiency and Compute Economics

Under standard baseline pricing, Carbon averages 123,000 output tokens per Intelligence Index task. This represents a 29% reduction in output generation volume compared to Claude Opus 5.5 (max), which averages 173,000 output tokens per task.

  • Claude Opus 5.5 (max): $10.04 per task (~23% more expensive than Carbon).
  • GPT-6 Astra (max): $6.25 per task (~1.2x cheaper than Carbon).
  • Gemini 4 Argon: $7.68 per task.

Carbon achieves a 9-point baseline score increase on the Intelligence Index over Argon while incurring virtually identical per-task inference costs ($7.76 versus $7.68).

Read the original article at artificialanalysis.ai.