Gemini 4 Carbon Launches: Google's New #1 Model Goes to Cyber Defenders First

What Happened: Carbon Arrives Two Weeks After Argon

On October 28, 2026, Google announced Gemini 4 Carbon, its newest frontier model and — by the company's own accounting — its strongest ever. The launch comes just two weeks after Gemini 4 Argon went worldwide, a cadence that itself tells the story: Google's researchers and models are now, in the company's words, "working hand in hand to advance capabilities faster," shipping a meaningful frontier upgrade in fifteen days rather than six months.

The announcement came from Koray Kavukcuoglu, the SVP now leading Google DeepMind after Demis Hassabis stepped aside. Carbon's headline positioning is cybersecurity and long-horizon tasks — complicated, multi-step jobs that run for hours or days — rather than casual chatbot conversation.

The most striking detail: you can't use it yet. Carbon is rolling out first to a subset of vetted cyber defenders through Google's Fairwind Program, with access phasing in gradually to additional trusted partners. Developers, businesses and consumers get it only after Google finishes iterating on guardrails — no public date given.

#1 on the Leaderboard — and It's Not Close on Cyber

Carbon debuts at the top of the newly updated v5.0 Artificial Analysis Intelligence Index with a score of 59, ahead of Claude Opus 5.5 and GPT-6 Astra. In Google's own comparison table, Carbon posts the best score on 15 of 17 benchmarks against Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1.

The cyber numbers are where Carbon separates from the field:

• SEC-bench Pro — 88.2% for discovering real-world vulnerabilities
• ExploitGym — 54.8% for turning a found vulnerability into a working exploit
• CWE-bench — 71.0% for patching flawed code

On internal benchmarks, Google says Carbon generates secure code at more than twice the success rate of Argon, cuts unsupported claims by 2x, and shows reduced misalignment and fabrication rates in live coding environments. It's also more resistant to prompt-injection attacks and less prone to taking shortcuts — the exact failure modes that have made agent security this year's hottest topic.

That capability profile is precisely why access is gated: a model that can autonomously find, exploit and patch critical vulnerabilities is enormously valuable to defenders — and dangerous in the wrong hands. It's the same reasoning that kept Argon's cyber tier behind Fairwind at launch.

Built for Tasks That Run for Days

The second pillar of the launch is endurance. Google says internal teams have watched Carbon "run for days to persist through dead ends to tackle complex, multi-step engineering tasks." Where most models degrade or wander over long sessions, Carbon is claimed to hold a thread — backtracking cleanly when a path fails instead of fabricating progress.

This matters beyond Google's walls. The industry's center of gravity has shifted from chat to agents that own entire workflows, and long-horizon reliability is the binding constraint on everything from codebase migrations to month-long research projects. If Carbon's endurance claims hold up outside Google, expect it to become the model behind the next wave of autonomous coding and research tools — the category where assistants like Cursor, GitHub Copilot and Gemini CLI compete on which underlying model best survives a twelve-hour shift.

Pricing and Access: Fairwind First, Everyone Else Later

When Carbon reaches the API, it will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. After the intro window, it steps up to $4/$20 — the same doubling pattern Argon is scheduled for after November 30.

Two access details stand out. First, Google says it is "actively engaged in the U.S. government's voluntary process for pre-release model testing" — a notable shift toward formal pre-launch evaluation at exactly the moment regulators on both sides of the Atlantic are scrutinizing frontier releases. Second, the Fairwind-first rollout means the first real-world Carbon deployments will be in security operations centers, not consumer apps. When it does land for everyone, expect it first across Google's own surfaces — the Gemini app, AI Mode in Search, and build tools like Google AI Studio and Vertex AI.

What Carbon Means for the AI Tool Market

For people choosing tools rather than following model launches, three practical takeaways:

1. The frontier is moving faster than tool integrations can absorb. Argon reached general availability on October 13; Carbon beats it two weeks later. No coding assistant, research tool or chat app re-benchmarks its defaults that quickly. The rational response is to prefer tools that let you swap models per task — multi-model routing is no longer a nice-to-have when the leaderboard changes every fortnight.

2. AI security tools are about to get a generational upgrade. Carbon's SEC-bench and ExploitGym leads will flow into vulnerability-scanning, code-review and SOC-assistance products within months. If you're evaluating security tooling this quarter, ask vendors a simple question: which model powers detection, and when does Carbon land in your pipeline? Meanwhile, the same capability means AI-generated attacks get better too — defenders adopting AI tooling is now an arms-race necessity, not an optimization.

3. "Gated launch" is becoming a marketing feature. Google is explicitly selling deliberateness — government-aligned pre-release testing, guardrail iteration before wide access — as a differentiator. After a year of rogue-agent incidents and safety-evaluator open letters, labs have learned that restraint at launch reads as seriousness. Expect the next Anthropic and OpenAI flagships to stage their rollouts the same way.

Frequently Asked Questions