Google Releases Gemini 4 Carbon to Cyber Defenders First, With No Date Set for Wider Access

Google released Gemini 4 Carbon on 28 October 2026, describing it as its strongest frontier model and making it available at first only to a subset of cybersecurity defenders through its Fairwind Program. The launch came two weeks after Gemini 4 Argon, and Google has set no date for general availability, according to Reuters.

The decision to open a flagship model to defenders before developers, businesses or consumers is the most consequential part of the announcement. It turns a benchmark race into a question about who gets access to offensive-capable software first.

What did Google announce?

Google DeepMind's Koray Kavukcuoglu wrote the launch post, and Google AI's account on X called Carbon "our strongest model with state-of-the-art cybersecurity and long horizon capabilities," as Runtime Wire reported. Reuters said Google claims state-of-the-art performance on vulnerability discovery, exploitation and software patching, and that on internal tests Carbon generates secure code at more than twice the success rate of Gemini 4 Argon.

Pricing is unchanged from Argon's introductory rate: $2 per million input tokens and $10 per million output tokens, according to Reuters and Constellation Research. Constellation read the release as evidence that Google is speeding up its frontier cadence, a framing the company's own timeline supports: two flagship-tier releases in 14 days.

TestingCatalog compiled the numbers Google published. Carbon scored 88.2% on SEC-bench Pro, which measures vulnerability discovery, 54.8% on ExploitGym, which measures exploitation, and 71.0% on Collinear's CWE-bench, which measures working fixes. TestingCatalog said these external cybersecurity evaluations were run without guardrails. It also reported that Google's Chrome team saw Carbon produce high-quality patches at more than 10 times the rate of another lab's large frontier model; the page does not name that lab.

Broader rankings are mixed. TestingCatalog listed Carbon first on the Artificial Analysis Intelligence Index v5.0 with a score of 59, and second on that firm's Coding Agent Index behind Claude Opus 5.5. It also listed the model first on Text Arena, second on Agent Arena and fourth on Code Arena's WebDev leaderboard. CNET, in a report carried by Yahoo Tech, said Carbon outperformed Gemini 4 Argon and OpenAI's GPT-6 Astra on benchmarks of overall capability, though CNET printed no scores. Runtime Wire noted that the Artificial Analysis index does not evaluate cybersecurity at all.

Why restrict access?

Google says the staged release reflects its participation in the US government's voluntary pre-release testing process, per TestingCatalog, and that wider availability will follow after the company collects feedback and iterates on guardrails. CNET quoted the same position: Google is working with the US government's review process and declined to say when broader access will open. CNET's own guess, based on past releases, was one to two weeks. That is CNET's estimate, not a Google commitment.

The reasoning is straightforward. A model that scores 54.8% on an exploitation benchmark without guardrails is useful to a defender triaging a codebase and equally useful to an attacker. Releasing it first to vetted defenders gives them a head start on patching before the same capability reaches the public. CNET tied the launch to recent AI-assisted discoveries of PHP vulnerabilities that affected a majority of the world's websites, which shows why defenders want the tool now.

Google did not describe who qualifies for Fairwind or how to apply. Reuters said the programme exists while safeguards continue to be tested, and one aggregator noted that applying does not guarantee access. Independent verification of the benchmark claims was not available in any of the coverage reviewed.

What does the safety record show?

Google's safety claims are specific. TestingCatalog reported that the safety stack carried over from Argon covers model training, real-time misuse defences, prompt-injection resistance and chain-of-thought monitoring. On Terminal-Bench 4.0 Google reported 0.0% successful cheating, with attempts at 0.4%. High-severity misalignment alerts in internal coding work fell by almost 80%, and attack success on Gray Swan's Indirect Prompt Injection Benchmark fell from 0.7% to 0.0%. Against Argon, Google reported roughly half as many unsupported claims and 26% less overreach.

All of those figures are Google's own and come from its launch materials. They should be read as a vendor's account until outside evaluators publish results, which cannot happen at scale while access is limited. That is the tension in the staged approach: the tighter the gate, the fewer independent eyes can test whether the gate is justified.

What happened to Genie 4?

Google separately introduced Genie 4, a world model that simulates physical environments, in the same week. CNET said it targets robotics and game design, adds audio, including background noise and dialogue, as a new capability, and is available only through a waitlist. It is a different product line from Carbon and carries less immediate security weight, but it shows Google pressing on two fronts: language models that act as engineers, and simulators that could train robots.

Why it matters for the lab race

Reuters framed Carbon as competition with OpenAI and Anthropic in coding and cybersecurity systems. CNET added that Google ended 2025 as a major AI player but has trailed peers, and that Argon was its first new frontier model in 11 months. Two releases in two weeks reverse that picture. Whether Carbon's lead holds depends on what outside benchmarks show once access widens.

Developers will care about a second consequence. Runtime Wire noted that Carbon's production performance is out of reach for most developers during the restricted rollout, so the headline price of $2 and $10 per million tokens is, for now, a price almost nobody can pay. The date general availability opens will reveal how confident Google is in its guardrails.

What we don't know yet

Several material facts are unconfirmed. Google has not given a general availability date. Reuters and Constellation did not publish benchmark methodology, and Fairwind's eligibility rules are undisclosed. No source reviewed offered an independent check of Google's cybersecurity scores or its safety statistics. CNET's one-to-two-week access estimate is speculation.

The next test is the first public evaluation of Carbon outside Google's own materials. If the exploit and patch numbers hold up, defenders gain a real tool. If they do not, the staged release will look more like caution than capability.