Gemini 4 Carbon: Google retakes first place in the Intelligence Index with a cybersecurity-focused model

Gemini 4 Carbon: Google reclaims first place on the Intelligence Index with a cybersecurity-focused model

🔎 A launch that smells like a power play

On October 28, 2026, Google DeepMind unveiled Gemini 4 Carbon, just fifteen days after the release of Gemini 4 Argon. Such a breakneck pace says a lot about the competitive pressure: OpenAI and Anthropic aren't letting up, and Google has decided to respond in a big way.

On paper, the result is unambiguous. Gemini 4 Carbon scores 59/100 on the Artificial Analysis Intelligence Index v5.0, ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57). Google thus reclaims the top spot in the rankings, which it had lost to Anthropic several weeks ago.

But what really sets Carbon apart isn't the score. It's the angle: for the first time, a frontier model is designed around cybersecurity as its primary use case. Vulnerability detection, controlled exploitation, automated patching — Google openly intends to become the go-to tool for defense teams. Provided you can get access, because the rollout is deliberately restrictive.

The Essentials

• 59/100 on the Artificial Analysis Intelligence Index v5.0: Gemini 4 Carbon reclaims first place ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57).
• A model built for cybersecurity: 88.2% on SEC-bench Pro (vulnerability detection), 54.8% on ExploitGym (exploitation), 71.0% on CWE-bench (patching).
• Strictly controlled access: the Fairwind Program reserves initial access for verified defenders (SOCs, CERTs, security researchers).
• ⚠️ Security results self-evaluated and without guardrails: Google measured the model's offensive capabilities by disabling its own protections. A methodological choice we detail below, and one that warrants your vigilance.
• Unchanged pricing: $2 / $10 per million tokens (input/output), identical to Argon — verify on ai.google.dev.

Recommended Tools

For an updated month-by-month comparison of these models, check out our ranking of the best LLMs of October 2026.

Gemini 4 Carbon, what exactly is it?

Direct answer: it's the third model of the Gemini 4 generation, a frontier model specialized in cybersecurity that posts the highest score ever recorded on the Intelligence Index v5.0 from Artificial Analysis.

Carbon is a direct continuation of Argon, released on October 13, 2026. Google kept the same base architecture and the same pricing — $2 per million input tokens, $10 for output — but redirected the post-training toward a massive corpus of vulnerability analysis, CVE reports, disarmed malicious code, and incident response scenarios.

The name "Carbon" is no accident. At Google, it evokes the structure of carbon: solid, versatile, capable of forming bonds with almost everything. The marketing message is clear — this model is meant to connect to every tool in a defense chain: SIEMs, scanners, sandboxes, CI/CD pipelines.

On the generalist front, Carbon retains Argon's strengths: extended context window, excellent performance in code and multi-step reasoning. This is not a niche model, but a generalist with a clear specialization.

The numbers that matter: what the benchmarks really say

Direct answer: Carbon dominates three cybersecurity benchmarks with significant margins, but these results are self-reported by Google and measured without guardrails — two essential details for reading these numbers correctly.

According to The Hacker News, the scores on SEC-bench Pro and CWE-bench exceed those of Claude Opus 5.5 and GPT-6 Astra by 6 to 9 points on the same tests. On ExploitGym, the gap is more modest but still in Google's favor.

The methodological point that cannot be ignored: VentureBeat points out that Google deliberately disabled the guardrails of its own model to measure its offensive capabilities. In other words, the 54.8% on ExploitGym represents the model's raw ceiling, not what an average user will get through the public API. This is an increasingly common practice among labs — OpenAI and Anthropic do the same in their risk evaluations — but it makes comparisons with models evaluated with protections tricky.

Second caveat: these three benchmarks are published by Google itself, not by an independent third party. Artificial Analysis, which publishes the Index v5.0, is independent and credible — so the 59/100 is solid. The cybersecurity scores, however, will have to wait for external validation. The Verge notes, moreover, that Google promises to open SEC-bench Pro to third-party evaluators "in the coming months".

Our take: the 59/100 on the Index v5.0 is a verifiable and remarkable fact. The cybersecurity scores are promising but should be taken with the appropriate caution until a third party has reproduced them.

The Fairwind Program: locked-down access, and that's by design

Straight answer: you won't be able to test Gemini 4 Carbon today, even if you pay. Google reserves initial access for members of the Fairwind Program, an onboarding program reserved for vetted defenders.

Concretely, the Fairwind Program targets:

• SOC and CERT teams (internal or commercial)
• Security researchers affiliated with recognized institutions
• Bug bounty hunters with a verifiable track record
• Defense tool vendors (scanners, EDR, threat intelligence)

The process includes identity verification, a contractual commitment regarding usage (defense only, no unauthorized offense) and request logging by Google. According to VentureBeat, Google has not announced a date for a general API opening, mentioning only a "gradual expansion."

Why such caution? The logic is defensible: a model trained to exploit vulnerabilities at 54.8% on ExploitGym is a double-edged sword. Releasing it unfiltered on the open market would hand attackers exactly the tool Google wants to put in the hands of defenders.

But there's a counterpoint that observers are pointing out: restricted access is also a way to escape immediate independent testing. It's impossible to reproduce the benchmarks if you can't query the model. The skepticism is healthy — and recent AI history has taught us that spectacular announcements with limited access always deserve a cautious wait before being judged.

In the meantime, if you're looking for a powerful model that's accessible right away, our comparison of the best free LLMs lists the options available without a paid signup.

Cybersecurity: Why Google Is Betting Everything on It

Direct answer: because it's the segment where demand is exploding, where training data is rare and valuable, and where Google has a structural advantage that neither OpenAI nor Anthropic can easily replicate.

Three factors explain this pivot:

1. The defense market is in a critical talent shortage. The gap between the number of cybersecurity professionals needed and the positions filled is measured in millions worldwide. A model capable of automating alert triage, code analysis, and patch writing addresses a massive economic need — and budgets that companies are ready to unlock.

2. Google owns unique training data. Years of vulnerability reports from Project Zero, the Google Cloud security infrastructure, data from VirusTotal, and the experience of Android facing billions of devices attacked daily. No competitor has this corpus.

3. It's a differentiator against Claude and GPT. On the general reasoning front, the three models are within 2 points of each other (59, 58, 57). Differentiation is therefore shifting to verticals. OpenAI bet on agents and screen control with GPT-6.1 Terra, Anthropic on code and enterprise. Google is choosing security — a territory where DeepMind's perceived expertise is a considerable marketing asset.

Worth noting: this vertical strategy is not without risk. If independent benchmarks don't confirm the claimed lead, Carbon risks being perceived as just another generalist model with security dressing. The Intelligence Index v5.0, tougher than previous versions, shows on the other hand that Google is progressing on the general foundation — and not just on specialization.

To dig deeper into the Google ecosystem, our article on Gemini Intelligence and Android as an intelligence system shows how Google orchestrates its models across its software empire.

Head-to-Head: Carbon vs Claude Opus 5.5 vs GPT-6 Astra

Direct answer: Carbon wins on the overall leaderboard and on security, Opus 5.5 remains the king of code and reliable agents, and GPT-6 Astra remains the most versatile for everyday use.

What the table doesn't show: availability has become a decision criterion just as important as performance. A model you can't access is of no use to you, no matter its score. Claude Opus 5.5 remains our default recommendation for development — our guide to the best LLMs for coding explains why.

On the agents front, where the gaps widen in real-world use, Carbon has not yet been independently evaluated to our knowledge. GPT-5.5 and Claude Opus have historically dominated this segment, and the October 2026 agentic leaderboard remains dominated by OpenAI. Our feature on the best LLMs for AI agents tracks this territory month by month.

Our position: Carbon is the most intelligent model of the moment on paper, but Opus 5.5 remains the most useful model for the majority of users, simply because it is accessible and proven.

What this changes for defense teams

Direct answer: if you work in a SOC or a CERT, Carbon is potentially the most significant tool since LLMs arrived in your field — but you'll first have to go through the Fairwind Program.

The most promising use cases:

• Automated alert triage: with 88.2% detection on SEC-bench Pro, Carbon could drastically reduce noise in SIEMs, where analysts still spend the majority of their time on false positives.
• Assisted patching: 71.0% on CWE-bench means the model generates a correct patch more than two times out of three on known vulnerabilities. That's a real time-saver for maintenance teams, even though human validation remains mandatory.
• Internal vulnerability research: use the model in "attack" mode on your own code, in a legal and controlled framework, before someone else does.

The point of concern: the 54.8% exploitation rate on ExploitGym, measured without guardrails, raises a governance question. Defense teams deploying Carbon will need to make sure that production protections are calibrated to prevent unauthorized offensive use — and Google has not yet published detailed documentation on the exact behavior of the public guardrails relative to the model's raw capabilities.

It is precisely this type of gap between "raw capability" and "exposed capability" that will need to be audited by third parties. We'll come back to this as soon as independent evaluations become available.

❌ Common Mistakes

Mistake 1: Taking the security scores at face value

What's wrong: the SEC-bench Pro, ExploitGym and CWE-bench scores are self-reported by Google and measured without guardrails. Taking them as-is into a purchasing decision means comparing apples and oranges with the scores of other models measured under different conditions.
The fix: wait for third-party evaluations, or test the model yourself via the Fairwind Program on your own use cases before committing.

Mistake 2: Believing you can access it right away

What's wrong: many articles announce the model as available. It is not available to the general public or via the standard API at this stage.
The fix: check the Fairwind Program eligibility criteria on the official Google blog, and in the meantime, use Gemini 4 Argon, which is available at the same price.

Mistake 3: Replacing a human analyst with the model

What's wrong: even at 88.2% detection, a 12% error rate on critical vulnerabilities is unacceptable without supervision. A patch generated with 71% success must be systematically reviewed.
The fix: treat Carbon as an amplifier for your analysts, not as a substitute. Human-in-the-loop workflows remain the norm in 2026.

Mistake 4: Overlooking accessible alternative models

What's wrong: focusing on the number 1 in the rankings while it's inaccessible, and ignoring excellent models that are actually available.
The fix: for research and document analysis, tools like those featured in our selection of the best LLMs for research do a remarkable job today already.

❓ Frequently Asked Questions

Is Gemini 4 Carbon available via the Gemini API?

Not yet for the general public. Initial access goes exclusively through the Fairwind Program, reserved for verified defenders (SOC, CERT, researchers). Google mentions a gradual expansion without a specific date. In the meantime, Gemini 4 Argon is available via the API at the same price.

How much does Gemini 4 Carbon cost?

The announced price is the same as Argon: $2 per million input tokens and $10 per million output tokens (prices announced in October 2026, check ai.google.dev). That's in the range of frontier models, significantly cheaper than some premium "reasoning" models.

Are the cybersecurity scores reliable?

They are self-reported by Google and measured without guardrails, so they should be considered a capability ceiling rather than real-world performance. Artificial Analysis' Index v5.0, on the other hand, is independent and credible. Security benchmarks will have to wait for third-party validation.

Can Carbon be used by attackers?

That's precisely the risk the Fairwind Program aims to limit: identity verification, contractual commitments, and request logging. But the history of AI shows that capabilities eventually spread. The debate over labs' responsibility regarding offensive capabilities is far from settled.

Should you ditch Claude or GPT for Carbon?

No, for two reasons: access is limited, and Carbon has not yet proven its superiority for everyday use cases (code, agents, writing). On paper, the three models are neck and neck. Follow our monthly comparison of the best LLMs for updates.

How do you integrate Carbon into a SOC?

Google presents Carbon as connectable to SIEMs, scanners, and CI/CD pipelines via its native toolchains. Integration details are not yet public for non-members of the Fairwind Program. Interested teams should follow announcements from Google Cloud Security.

✅ Conclusion

Gemini 4 Carbon takes back the top spot on the Intelligence Index with a well-deserved 59/100 and a clear strategy: making cybersecurity Google's differentiating battleground against Anthropic and OpenAI. The security numbers are impressive but self-reported and without guardrails — caution is still warranted until third parties have validated them. If you work in defense, apply to the Fairwind Program; for everyone else, Gemini 4 Argon and Claude Opus 5.5 remain the pragmatic choices of the moment. And to make sure you don't miss any upcoming releases, our monthly ranking of the best LLMs is updated every month.

📚 Sources

• Official announcement of Gemini 4 Carbon — Google Blog
• Artificial Analysis — Intelligence Index v5.0
• Google unveils Gemini 4 Carbon — The Hacker News, October 2026
• Gemini 4 Carbon reclaims the #1 spot on the ranking — VentureBeat, October 2026
• Google Gemini 4 Carbon and cybersecurity — The Verge, October 2026
• Google Cloud Security
• Project Zero — Google
• VirusTotal

Note: the cybersecurity scores (SEC-bench Pro, ExploitGym, CWE-bench) come from Google's self-evaluations published in October 2026 and have not yet been validated by an independent third party.
```