Google DeepMind announces Gemini 4 Carbon: AA Intelligence Index v5 #1; SOTA cyber find/exploit/patch benches; Fairwind limited rollout after Argon

Why it matters

【Key points】

On 2026-10-28 Google DeepMind/Google announced Gemini 4 Carbon after Gemini 4 Argon (~2 weeks prior), emphasizing long-horizon agentic engineering and cyber-defense lifecycle (find → exploit-validate → patch). Vendor cites: Artificial Analysis Intelligence Index v5.0 score 59 (#1); SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%; WebDev Arena +89 Elo to #4 (1768). Intro price matches Argon ($2/$10 per M in/out; cached input 95% off; then $4/$20). Rollout: cyber-defender subset → Fairwind partners; U.S. voluntary pre-release testing; broader surfaces “soon,” not GA.

【Verifiable】

- DeepMind: https://deepmind.google/blog/gemini-4-carbon-our-next-frontier-model/ (RSS 2026-10-28T17:00:58Z)

- Google Blog: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ (17:00:00Z)

- Vendor cites (not re-run): AA v5 #1@59; cyber benches above; >2× secure-code vs Argon; Chrome >10× patch quality vs another lab frontier; Gray Swan IPI 0.0% (Argon 0.7%); Terminal-Bench 4.0 cheat success 0.0%.

- Peers: e/843 Argon global API; e/691 Argon trusted-cyber start; e/845 Argon Safety Report; e/1059 GPT-6.1 Terra; e/1050 Mythos/Fable.

【Our read】

Background: Argon→Carbon in two weeks compresses same-gen iteration toward long-horizon engineering + cyber lifecycle with Fairwind/gov pre-release restraint.

Peers: vs GPT-6.1 Terra/Astra, Claude Opus-class, Fable/Mythos, Carbon bids “AA #1 + full-stack defense” at Argon intro price. Scores are mostly vendor-cited / single-harness (press notes possible Antigravity CLI)—don’t claim universal dominance. Invite-only trails open competitor APIs.

Impact: near-term for invited defenders (SOC, vuln/patch). Most developers/Chat users get preview + price anchor until GA. Later GA would pressure OpenAI/Anthropic on agentic coding and security tooling. Risks: biweekly eval noise; alignment claims must hold. Carbon model+rollout only—not Gemini App immersives or XREAL.