Google's Gemini 4 Carbon went out on Wednesday to a small circle of vetted cyber defenders, and almost everyone else has to wait. Google announced the model on October 28, 2026, two weeks after Gemini 4 Argon, and called it its strongest frontier model for cybersecurity and long-horizon work. The company has not set a date for general availability.
Carbon is not in the Gemini app, AI Mode in Search or the public Gemini API as of launch. Technology.org described the launch as Google's strongest model for cybersecurity work, and SecurityWeek reported the company's results for vulnerability discovery, exploitation and patching. American security teams and developers can read about the model but cannot use it yet.
Google calls this a defender-first approach. The same capabilities that find and fix software flaws can also be misused, so the company is starting with partners it can vet. Google's launch announcement said Carbon will first reach "a subset of cyber defenders" and then roll out gradually to more trusted partners, as FelloAI quoted from the blog post.
How to get access to Gemini 4 Carbon through Fairwind
Access starts this week with a small group of cyber defenders. Google will widen it in stages to more partners in its Fairwind Program, which the company says has more than 650 participating organizations, before developers, businesses and consumers get the model, according to SecurityWeek and Creators Toolbox.
Fairwind covers governments, critical infrastructure operators and major technology platforms. Google DeepMind runs background checks on organizations that apply, so individuals cannot join directly. FelloAI reports that security teams at governments, critical infrastructure operators, hospitals, telecoms and energy firms are the first group that can get Carbon, and only if Google admits them through Fairwind.
The contrast with Argon is sharp. Argon is already in the Gemini app for paid subscribers, and some Fairwind partners already had exclusive access to its cyber capabilities. Carbon's first phase is narrower than Argon's rollout, which signals how much more caution Google is applying to this model. For a sense of how seriously AI labs now treat misuse risk, see our coverage of Anthropic's IPO risk factors.
Gemini 4 Carbon benchmarks compared with GPT-6 Astra
Google's headline numbers are its own. The company ran its security benchmarks without guardrails, and public versions with safety filters may score differently.
| Test | Carbon score | What it measures |
|---|---|---|
| SEC-bench Pro | 88.2% | Vulnerability discovery |
| ExploitGym | 54.8% | Exploitation |
| CWE-bench | 71.0% | Patching |
On SEC-bench Pro, Carbon scored 88.2% against 85.4% for OpenAI's GPT-6 Astra, according to BigGo Finance. That gap is small, and it rests on one vendor's test suite.
On the Artificial Analysis Intelligence Index version 5.0, Carbon scored 59, one point ahead of Claude Opus 5.5. Vals AI ranked Carbon second of 29 models with 70.36% on its Vals Index. On WebDev Arena, Carbon posted 1769 Elo for fourth place, a gain of 89 Elo over Argon.
The hallucination rate that complicates the rankings
The independent number that stands out is not a benchmark score. WinBuzzer reported that Artificial Analysis measured Carbon's hallucination rate on its factual-knowledge test at 11%, up from 4% for Argon, even though Carbon answered more questions correctly.
That tradeoff matters for security work. A model that answers more questions can also produce more confident wrong answers, and an analyst who acts on a fabricated detail about a vulnerability can lose hours or misdirect a patch. So far, the figure rests on a single independent measurement, and the benchmark results may shift once broader testing arrives.
Gemini 4 Carbon price and what is still unclear
The pricing has two tiers, and only the lower one is in effect at launch:
- Introductory API price: $2 per million input tokens and $10 per million output tokens, with cached input 95% off, per Crypto Briefing and Anadolu Agency.
- Standard price after the introductory period: $4 per million input tokens and $20 per million output tokens, per NeuralTrust.
- Artificial Analysis put the cost per Intelligence Index task at standard pricing at $7.76, using 123,000 output tokens, according to SaaS City.
Google has not set an end date for the introductory rate. SaaS City lists $4 and $20 as the standard rate with a 50% launch discount. One source table lists November 30, 2026, as the end of Argon's introductory period, but that date is unconfirmed. Teams that budget for Carbon should treat the introductory numbers as provisional.
For context on how OpenAI has moved prices for agent workloads, see our report on GPT-6.1 Terra cutting API prices in half for computer-use agents.
Federal pre-release testing and the state rules around it
Google says it is taking part in the U.S. government's voluntary pre-release testing before wider release. The sources reviewed do not describe what that testing involves or how long it takes. One source notes that participation does not amount to government certification, according to Kingy AI.
The federal backdrop is thin. Congress has not enacted a single federal AI law despite more than 150 federal AI bills introduced over the past two sessions, according to a 2026 report from the Transparency Coalition summarized by TechTimes. The Trump administration pushed to preempt state AI rules, and that effort collapsed in the Senate on a 99-1 vote in July 2025.
States are moving on their own. Virginia Gov. Abigail Spanberger signed an executive directive on October 23 requiring a statewide plan for cyberattacks that involve AI, as 1023 Jack reported.
What to do if you run a security team or build on AI
If you run a security team, check whether your organization qualifies for Fairwind, and watch for access announcements from Google. Fairwind covers governments, critical infrastructure operators and major technology platforms, so smaller firms may not qualify.
If you are a developer, plan your budget at standard pricing of $4 and $20 per million tokens, not the introductory rate. Test any benchmark claim on your own workloads, and pay particular attention to factual accuracy given the hallucination figure.
Keep your current model access in place for now. Carbon is not in the public Gemini API, and no general release date has been announced.
What comes next for Carbon
Three signals will matter in the coming weeks. The first is how quickly Google widens Fairwind beyond its initial group of defenders. The second is whether a second independent lab confirms the rise in hallucinations. The third is whether Google publishes an end date for the introductory price. Together they will show whether Carbon's top ranking holds outside Google's own tests, and how far the company is willing to open a model it considers too useful to hold back entirely.
