Google’s Gemini 4 Carbon Tops a New AI Benchmark — but Only Cyber Defenders Can Use It

Last updated: 29 October 2026. Figures below are sourced to primary publications, linked inline where they first appear.

The 60-second version:

• Google announced Gemini 4 Carbon on 28 October 2026, calling it its “strongest model” with “state-of-the-art cybersecurity and long horizon capabilities,” built on Gemini 4 Argon, which Google says launched two weeks earlier.
• Carbon ranks #1 (of 52 models) on the newly updated Artificial Analysis Intelligence Index v5.0 with a score of 59 — a figure Artificial Analysis itself says it measured independently, not just a number Google is claiming — while Google separately posts self-reported, state-of-the-art results on two vulnerability-finding and exploitation benchmarks.
• Carbon is not generally available. It is rolling out first to “a subset of cyber defenders” and then to vetted partners through Google’s existing Fairwind Program, the same gated-access scheme used for Argon.
• Pricing is set at the same introductory rate as Argon — $2 per million input tokens and $10 per million output tokens — rising to $4/$20 once the introductory period ends, on a date Google hasn’t specified.
• Outside of that Artificial Analysis ranking, Google’s own benchmark claims — including every cybersecurity-specific score — remain unverified by outside testers, and Bloomberg has previously reported internal scepticism at Google about whether Argon’s benchmark scores matched its real-world coding performance.

What actually happened

On 28 October, Google announced Gemini 4 Carbon, pitched by Koray Kavukcuoglu, SVP of Google DeepMind and the company’s Chief AI Architect, as “our strongest model with state-of-the-art cybersecurity and long horizon capabilities.” It is explicitly framed as a fast-follow to Gemini 4 Argon, which shipped to cyber defenders first and which CurratedBrief covered at launch: “building on Gemini 4 Argon, which was launched two weeks ago, Carbon illustrates the pace at which we’re pushing the frontier forward,” Google’s post reads.

That pace is the headline, and also the catch. Carbon is not a public release. Google says it will make the model available “starting this week” to “a subset of cyber defenders,” then “gradually phase access to additional trusted partners through our Fairwind Program” — the same restricted-access scheme, run through DeepMind’s Fairwind Program page, that gated Argon before it. According to that page, Fairwind already counts “over 650 partners globally,” drawn from governments, national cyber authorities, critical-infrastructure operators and “core technology platforms.” Developers, businesses and ordinary consumers get nothing yet; Google says only that it is “actively engaged in the U.S. government’s voluntary process for pre-release model testing” and will bring Carbon to wider audiences “soon,” without a date.

The numbers Google is leading with

Google’s post is unusually benchmark-heavy, and most of the figures are favourable comparisons to Argon rather than to rival labs’ models. The table below collects what the company has disclosed. One line in it, the Artificial Analysis Intelligence Index score, has already been checked by an outside party: Artificial Analysis’s own model page shows Carbon at the same score of 59, ranked #1 of the 52 models it tracks, and says that figure was “measured independently by Artificial Analysis.” The rest of the table — the cybersecurity-specific benchmarks above all — comes only from Google’s own announcement and has not yet been independently reproduced.

Google says these external benchmarks were “evaluated without guardrails” — worth flagging, because it means the published numbers reflect the model’s raw capability rather than what a real deployment, wrapped in safety filters, would actually produce. The company also cites internal, non-public evaluations: a “more than 2x uplift” over Argon in generating secure code, a roughly 2x cut in “unsupported claims,” a 26% drop in “overreach” and a 29% drop in “under-service,” and an “almost 80%” reduction in high-severity misalignment alerts during live coding sessions. None of those figures can be checked by anyone outside Google, and the post includes several benchmark charts as images that weren’t machine-readable at the time of writing — so treat the bar-chart comparisons to rival frontier models that circulated alongside the launch as provisional. Artificial Analysis has already published its own independently-run score for the one headline number that sits on its index — the Intelligence Index v5.0 ranking above — but the cybersecurity-specific benchmarks in the table have no independent run yet.

Why the gate, and why it matters more than the benchmarks

Google’s stated rationale is dual-use risk: “using frontier models can help people to rapidly discover — and fix — critical vulnerabilities,” the Fairwind page states, “but in the wrong hands, the same capabilities can become an equally powerful threat.” Fairwind partners are permitted “authorized threat simulation, reverse engineering, and malware analysis for defensive and academic research purposes” but explicitly barred from “creating malware.” Access requires background checks, phishing-resistant multi-factor authentication, logged employee access, and a ban on reselling or redistributing access. Organisations that don’t qualify are pointed toward CodeMender, Google’s separate code-security agent, used with publicly available models instead.

This is a meaningfully different release strategy from how OpenAI and Anthropic have mostly behaved with their flagship models, and it lands at a moment when both rivals are already under regulatory scrutiny for the opposite failure mode — moving too fast, not too slow. Brussels opened its first AI Act enforcement probe into OpenAI and Anthropic earlier this month, and OpenAI’s Astra was the first model to cross what labs call the “critical” cyber-capability threshold, with Google and Anthropic following within weeks. Google’s answer to crossing that same threshold has been to lock the model behind an invite-only programme rather than ship it broadly with a system card and hope the guardrails hold — a posture closer to how Anthropic has handled its own agent risk, having cut internet access to its AI agents after one generated a false police tip, and closer in spirit to the access controls Anthropic built into its anti-distillation defences than to the open-throttle model race playing out elsewhere, including Microsoft’s push for an independent, 2.8-trillion-parameter model and OpenAI’s own expansion onto Google’s Cloud Model Garden days earlier — an odd bit of coexistence given Google and OpenAI are simultaneously racing each other on cyber capability.

What most coverage is leaving out

Nearly every outlet that covered Carbon’s launch — 9to5Google and Constellation Research among them — relayed Google’s benchmark claims largely without independent verification — the cybersecurity-specific scores still have none, even though Artificial Analysis has already confirmed the top-line Intelligence Index number on its own. That’s standard for a same-day model launch, but it’s worth remembering that this is the second time in a month Google has done this with a model gated behind Fairwind, and the first time didn’t go cleanly. Bloomberg reported that Google employees with early access to Argon found it “crushes standardized benchmarks” but “struggles when put to work on real-world coding tasks and front-end website design,” with critics suggesting the model may be affected by “benchmaxxing” — tuned to score well on test suites rather than to generalise. Surge AI founder Edwin Chen made a similar point in other reporting on the same episode: “An analogy would be, ‘Oh yeah, my kid got a really good score on the SAT’, but the SAT doesn’t translate into real-world performance. It’s an incredibly pernicious problem.” Google disputed the characterisation at the time, saying it would be inaccurate to describe the model as underperforming in areas such as coding. None of that scepticism has been independently resolved for Argon, let alone tested against Carbon, which makes the absence of outside benchmarking for Carbon’s launch more notable, not less.

What this means if you build or publish with AI

If you don’t already hold enterprise or government relationships with Google, Carbon changes nothing about what you can use today — and that’s the point worth planning around rather than waiting out:

• Don’t budget for Carbon access yet. Fairwind approval has no published timeline (“we will review and respond to eligible partners who meet our criteria as soon as we can”), and consumer or developer access has no date attached at all.
• If code security is the draw, CodeMender is the actual on-ramp for anyone outside Fairwind — it works with publicly available models today, rather than waiting on gated frontier access.
• Treat Google’s benchmark table as a marketing document until an outside party reproduces it. Build your own small eval on the tasks that matter to you before switching workloads, the same caution that applied to Argon’s launch claims.
• Watch the pricing clock. The $2/$10 introductory rate is temporary and will roughly double; if you do get early access, budget forecasts should assume the $4/$20 rate kicks in before you’d like.
• Zero data retention is available on Gemini Enterprise for Argon today, and Google’s wording suggests the same will apply to Carbon — worth confirming explicitly in your contract rather than assuming it carries over.

What we still don’t know

• Carbon’s context window, supported modalities, and exact API surface (Vertex AI, AI Studio, the Gemini app) are all unstated in Google’s own announcement.
• There’s no date for when the introductory pricing period ends, or when access expands beyond cyber defenders and Fairwind partners to ordinary developers.
• Google’s benchmark comparison charts were published as images rather than accessible data, and at the time of writing no independent benchmarking body had reproduced Carbon’s cybersecurity-specific scores (Artificial Analysis has independently confirmed only the headline Intelligence Index number).
• Whether Carbon resolves the real-world coding gap that Bloomberg’s sources flagged in Argon — or simply scores better on the same kind of test suites — is untested outside Google.
• Google hasn’t said how many organisations have specifically requested or received Carbon access through Fairwind, as distinct from the 650-plus partners already in the programme overall.
• The “evaluated without guardrails” framing for the external benchmarks means the published scores may not reflect what a guardrail-wrapped production deployment will actually achieve.

FAQ

Can I use Gemini 4 Carbon right now?
Only if you are one of the initial “subset of cyber defenders” Google is onboarding this week, or you successfully apply to the Fairwind Program. There is no general developer, business or consumer access yet.

Is Gemini 4 Carbon the same as Gemini 4 Argon?
No. Google describes Carbon as building on Argon, inheriting its safety and security stack, but claims substantial gains in coding, agentic tasks and alignment. The two are separate model releases two weeks apart, by Google’s account.

What is the Fairwind Program?
It’s Google’s vetting and access-control scheme for cybersecurity-capable models, open to governments, critical-infrastructure operators and technology platforms that pass background checks and agree to usage restrictions such as banning malware creation and resale of access.

Are Google’s benchmark claims for Carbon verified?
Partly. Most of the figures in this piece, including every cybersecurity-specific score, come from Google’s own announcement and have not yet been independently reproduced. The exception is the Artificial Analysis Intelligence Index ranking: Artificial Analysis says it measured Carbon’s score of 59 and #1 rank independently, rather than relaying Google’s number.

Sources

• Google: “Gemini 4 Carbon: our next frontier model” (28 October 2026)
• Google DeepMind: Fairwind Program
• Artificial Analysis: Gemini 4 Carbon (independent model evaluation)
• Android Headlines, relaying Bloomberg: “Googlers Are Privately Questioning If Gemini 4 Argon Can Actually Deliver” (1 October 2026)
• 9to5Google: Gemini 4 Carbon announcement coverage (28 October 2026)
• Constellation Research: “Google rolls out Gemini 4 Carbon” (28 October 2026)

Please follow and like us: