Fifteen days after Gemini 4 Argon reached general availability, Google has a successor in the hands of security teams. Gemini 4 Carbon, announced Wednesday by Google DeepMind, ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59 (Google, 2026). Google built the launch around cyber defense and multi-day engineering work, and Sundar Pichai used his LinkedIn post on the release to point back to his earlier promise of periodic model releases.

A subset of cyber defenders gets Carbon this week. Google will then add trusted partners through its Fairwind Program and bring the model to developers, businesses and consumers once it finishes gathering feedback on guardrails, with no date given for that step. Carbon launches at Argon's introductory price of $2 per million input tokens and $10 per million output tokens, which rises to $4 and $20 when the introductory period ends (Google, 2026).

Carbon's Exploitation Score Matters Most to Your Patch Queue

Google ran its external cybersecurity benchmarks with guardrails off. Carbon scored 88.2% on SEC-bench Pro, which tests whether a model can uncover flaws in real-world software, and 71.0% on CWE-bench, which tests whether it can write a working fix at the root cause (Google, 2026).

The 54.8% on ExploitGym deserves the closest read (Google, 2026). That benchmark measures whether a model can prove a flaw is exploitable, and security teams use that answer to decide which flaws to patch first. Google says defenders increasingly rely on models for this check. A model that confirms exploitability can shorten triage, and the same capability is why Google limits who gets its strongest cyber models.

On internal benchmarks, Google reports Carbon more than doubles Argon's success rate at generating secure code, and its Chrome team found Carbon more than 10 times better than another lab's large frontier model at producing high-quality patches (Google, 2026). Wiz, now part of Google Cloud, observed Carbon leading on web penetration testing benchmarks. Carbon also blocked every attack on Gray Swan's indirect prompt injection benchmark, a 0.0% success rate compared with Argon's 0.7% (Google, 2026), which matters to any team running agents that read tickets, email or web pages containing hidden instructions.

Fairwind Membership Decides Who Tests Carbon First

Google DeepMind runs Fairwind as a vetted early-access program with more than 650 partners worldwide (Google DeepMind, 2026). Priority goes to these groups:

• Governments and national cyber authorities
• Critical infrastructure operators in healthcare, telecommunications, energy and financial networks
• Core technology platforms whose software sits underneath millions of downstream users

Partners agree to use phishing-resistant multi-factor authentication, to limit model access to internal cybersecurity, incident response and penetration testing teams, and to track which employees use it. You apply through your Google Cloud account team, and Google runs background checks on each applicant organization. Organizations that do not qualify can use Google's CodeMender security agent with publicly available models.

Google announced Argon to Fairwind partners on September 30 and opened it to developers and enterprises through its API, and to Google AI Pro and Ultra subscribers, on October 13 (Google, 2026). The same October 13 post said Google had begun adding hundreds more Fairwind partners, so the program expands alongside each release.

A Two-Week Release Pace Tests Your Model Approval Process

Pichai described Carbon as evidence of how fast Google can move, and the dates support his claim: Argon reached general availability on October 13, and Carbon followed on October 28. If your architecture review board takes longer than 15 days to approve a model, a stronger Google model will arrive before the review of the last one finishes. Teams that approved Argon this month will need a faster path for successor models from a vendor they have already vetted.

Budgets need the same attention. Argon's introductory rate ends November 30, after which prices double to $4 per million input tokens and $20 per million output tokens (Google, 2026). Carbon starts at that same introductory rate, and Google has not said when Carbon's introductory period ends, so any forecast built on launch pricing for either model should carry the higher rate as well.

Fairwind Partners Kept the Unguarded Argon After the Public Release

One difference between Fairwind partners and everyone else outlasted general availability. Google's October 13 Argon post, published the day the model opened to the public, invited defenders to apply to Fairwind for access to Argon without cyber guardrails. The same post says the public Argon declines harmful requests, including requests to write exploit code, more reliably than earlier Gemini models.

Carbon's headline security scores come from that guardrails-off configuration. Google has not said whether Carbon will follow the Argon policy, or how much of the benchmark performance its public version will retain once misuse defenses are in place.

Google's caution has a clear rationale, because exploitation capability serves a red team and an attacker equally well, and Google says it is taking part in the U.S. government's voluntary pre-release testing process for Carbon. Your planning should account for the split anyway. Unless your organization qualifies for Fairwind, your penetration testers will likely work with a guarded Carbon whose security performance Google has not published.

Eligibility for a vendor's early-access program now sets the ceiling on your red team's tooling.

I hope Google publishes a schedule for moving Fairwind models to general release, along with benchmark results for the guarded public versions, because security teams can build a testing calendar around firm dates and published numbers.

What Happens Next

• Will Carbon reach general availability on Argon's two-week timeline, or will Google hold it in Fairwind longer because of its higher exploitation scores?
• Will Google publish benchmark results for the guarded public version alongside the Fairwind numbers?