AI Research

Google Gemini 4 Carbon Now Available to Cyber Defenders — #1 on Intelligence Index at 59, SEC-bench 88.2%, Fairwind Program Access

Google Gemini 4 Carbon model launch with Intelligence Index #1 ranking and cybersecurity benchmark scores

Google Gemini 4 Carbon: #1 on Intelligence Index at 59, Cybersecurity Crown, Fairwind Program Access

Google launched Gemini 4 Carbon on October 28, 2026 — its strongest frontier model to date, claiming the #1 spot on Artificial Analysis Intelligence Index v5.0 at 59 (edging out Claude Opus 5.5 at 58 and GPT-6 Astra at 57). Carbon delivers state-of-the-art cybersecurity capabilities across the full defense lifecycle: vulnerability discovery (SEC-bench Pro 88.2%), exploitation validation (ExploitGym 54.8%), and patching (CWE-bench 71.0%).

Access is initially restricted to the Fairwind Program — Google's curated cyber defender ecosystem with 650+ partners (governments, critical infrastructure operators, core tech platforms). Broader availability via API, AI Studio, Gemini app, and Google Antigravity began November 10 for Google AI Ultra subscribers, with general developer access "as soon as possible."

Benchmark Leadership: Intelligence Index, Cybersecurity, Coding

Benchmark Carbon Opus 5.5 GPT-6 Astra Argon Carbon Advantage
Artificial Analysis Intelligence Index v5.0 59 (#1) 58 57 50 +9 vs Argon
SEC-bench Pro (vuln discovery) 88.2% — — 83.4% +4.8 pp
ExploitGym (exploitation) 54.8% — — 42.0% +12.8 pp
CWE-bench (patching) 71.0% 67% (1-attempt) 68% 68.0% +3.0 pp
Terminal-Bench 4.0 (AA) 65.2% 56% (Opus max) — 57.1% +8.1 pp
FrontierSWE 2 61.2% 62.3% 65.5% 55.0% Trails Astra
OSWorld-2.1 offline 81.5% 87.2% — 74.7% Trails Opus
Agents' Last Exam v1.1 77.4% — — 69.0% +8.4 pp
GraphWalks 96.2% — — 84.2% +12.0 pp
Terminal-Bench-Science 64.8% — — 52.9% +11.9 pp
AutomationBench 53.9% — — 51.3% +2.6 pp

Critical caveat: Google states cybersecurity evaluations were run "without guardrails." The restricted version customers receive may score differently. Independent four-attempt CWE-bench (Artificial Analysis) shows Grok 4.7 at 81%, Opus 5.5 at 79% — Carbon's four-attempt result not yet published.

The Three-Stage Cybersecurity Defense Lifecycle

Google organizes Carbon's security results around the operational cyber defense workflow:

Stage Benchmark Carbon Score Operational Meaning
1. Discover SEC-bench Pro 88.2% Autonomously uncover complex flaws in real-world software
2. Validate ExploitGym 54.8% Check which vulnerabilities are actually exploitable
3. Remediate CWE-bench 71.0% Turn validated findings into working root-cause patches

Internal validation: Google reports >2× Argon's success rate for generating secure code. The Chrome team observed >10× advantage over a competing lab's model (unnamed) in producing quality patches. Wiz (Google Cloud security platform) confirms Carbon leads on web penetration testing benchmarks for their Scan for Good initiative.

Fairwind Program: Selective Access for Cyber Defenders

Requirement Detail
Partners 650+ organizations
Priority tiers Governments/national cyber authorities → Critical infrastructure → Core tech platforms
Background checks Mandatory for all partner personnel
Authentication Phishing-resistant MFA required
Access controls Restricted to internal security/incident-response/pen-test teams
No sharing Partners may not share or resell access
Rollout phases Subset of cyber defenders (Oct 28) → Gradual expansion to approved partners → Public availability

Precedent: Argon launched for Fairwind partners Sep 30 → Public Oct 13 (13 days). Carbon announced Oct 28 → Ultra subscribers Nov 10 (13 days). Full public timeline not specified — "as soon as possible" after guardrail iteration.

Pricing: Introductory Rates, Then 2×

Tier Input (per 1M) Output (per 1M) Cached Input Status
Introductory (until Nov 30) $2 $10 95% off ($0.10) Active
Standard (post-intro) $4 $20 95% off ($0.20) After Nov 30
Argon (current) $2 $10 95% off Stable
Fable 5.2 (Anthropic) $8 $40 $0.20 Stable
GPT-6 Astra (OpenAI) ~$10 ~$50 ~$1.00 Stable

Cost-effectiveness analysis: Carbon at introductory pricing is 2× cheaper than Fable 5.2 and 5× cheaper than Astra on standard rates. But at standard $4/$20, Carbon matches Fable 5.2's fast mode price ($16/$80 for 2.5× speed) for 1× speed. The 15× cost premium vs. Sol ($1.07/test per MathArena) for marginal gains is a key consideration.

The Three-Way Cybersecurity AI Race

Dimension Gemini 4 Carbon (Google) Claude Mythos 5.2 (Anthropic) OpenAI Security
Access Model Fairwind Program (650+ partners) Verification Programs (Cyber/Life Sciences) Opt-in watermarking, Dots agents
Benchmark Lead SEC-bench, ExploitGym, CWE-bench Terminal-Bench, FrontierCode, MathArena Computer-use, browser automation
Pricing (Intro) $2/$10 $8/$40 ~$10/$50
Pricing (Standard) $4/$20 $8/$40 ~$10/$50
Availability Ultra subscribers Nov 10; Fairwind phased General (Fable) / Vetted (Mythos) General
Safety Model Guardrails iterated post-Fairwind feedback Tiered: Fable (strong) / Mythos (relaxed) Watermarking, opt-in, Dots approvals
Enterprise Integration Google Antigravity, CodeMender, Wiz Enterprise Frontier Safeguards, ZDR, Claude Security Dots, Wallet, Codex, Enterprise

Google's positioning: Carbon as the "long-horizon agentic work" specialist — multi-step coding, debugging, open-ended scientific research using more tokens for thorough exploration, deep reasoning, extended verification loops. Argon remains the daily driver for most software engineering.

MathArena Leadership — At 15× Cost

Carbon scores 85.4% (±3.0) on MathArena vs. Fable 5.2 at 82.8% and GPT-6.1 Sol at 80.6%. But per-test cost: Carbon $15.70 vs. Sol $1.07 (15×). For math-heavy workloads, the ROI depends on whether the 2.6–4.8 pp accuracy gain justifies the compute spend.

Safety: Prompt Injection Resistance, Anti-Cheating

  • Gray Swan indirect prompt injection: 0% success rate vs. Argon 0.7%
  • Terminal-Bench 4.0 cheating attempts: 0% succeeded (attempts fell to 0.4%)
  • US government voluntary pre-release testing: Google actively participating

Strategic Context: Four Models in One Year

Carbon is Google's fourth Gemini 4 family model in ~2 months:

  1. Gemini 4 (base) — Foundation
  2. Gemini 4 Argon (Sep 30/Oct 13) — Daily driver, $2/$10
  3. Gemini 4 Carbon (Oct 28/Nov 10) — Frontier cyber/long-horizon, $2/$10 intro
  4. Gemini 4 Hydrogen (Nov 12) — Volume tier, $1.50/$7.50 (P6 Story 13)

This rapid iteration cadence (Argon → Carbon in 15 days) mirrors Anthropic's Opus/Sonnet/Haiku tiering but compressed. The Fairwind selective rollout for Carbon mirrors Anthropic's Verification Programs for Mythos — both labs converging on tiered capability deployment for dangerous domains (cyber/bio).

What Comes Next

  1. Fairwind expansion — Gradual partner onboarding; feedback drives guardrail iteration
  2. Public API launch — Timeline TBD; "as soon as possible" after safety validation
  3. Four-attempt CWE-bench — Independent verification needed (Artificial Analysis)
  4. Hydrogen volume adoption — $1.50/$7.50 tier competes with Haiku 5.5, Sonnet 5.5, Luna
  5. OpenAI/Anthropic response — Expected Astra price adjustment; Mythos 5.2 already launched
  6. Enterprise security adoption — Wiz integration, CodeMender, Antigravity as wedge

The Bigger Picture

Gemini 4 Carbon represents Google's bid for the cybersecurity AI crown — a domain where benchmarks translate directly to revenue (vulnerability management, pen-testing, secure code generation). The Fairwind Program creates a moat of trusted relationships with governments and critical infrastructure.

But the guardrail gap (eval without guardrails vs. deployed with guardrails), coding benchmark gaps (trails Astra on FrontierSWE, Opus on OSWorld), and 15× math cost mean Carbon is a specialist tool, not a universal upgrade. Argon remains the pragmatic choice for most developers; Carbon is for when the stakes — and the token budget — justify the frontier.


Related AIPress Coverage

Sources

  • Google Blog (Oct 28, 2026): "Introducing Gemini 4 Carbon" (Koray Kavukcuoglu)
  • Google Blog (Nov 10, 2026): "Gemini 4 Carbon is now available"
  • Artificial Analysis (Oct 28, 2026): "Gemini 4 Carbon is the new #1 on Intelligence Index v5"
  • The Verge (Oct 28, 2026): "Google launches Gemini 4 Carbon with 'significant leap' in cybersecurity"
  • VentureBeat (Oct 28, 2026): Detailed benchmark table, Fairwind analysis
  • SecurityWeek (Oct 28, 2026): "Google Touts Powerful Cybersecurity Capabilities in Gemini 4 Carbon"
  • Digital Trends (Oct 28, 2026): "Google's Gemini 4 Carbon is built for complex coding and cybersecurity tasks"
  • The Superintelligence News (Oct 2026): Independent claim-check, four-attempt CWE-bench gap
  • WinBuzzer (Oct 2026): Comparative benchmark analysis
  • Google DeepMind / Artificial Analysis leaderboard data (Oct–Nov 2026)

Building something with AI?

DevsIsle designs and ships AI systems, agents and integrations for teams that need it done properly.

Talk to our team →