Google’s Gemini 4 Carbon Tops the AI Index at 59 — and Only Cyber Defenders Can Use It Yet
Key takeaway
Google DeepMind announced Gemini 4 Carbon on October 28, 2026, calling it its strongest model yet for cybersecurity and long-horizon engineering work. It ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59, but you cannot use it yet: access starts with vetted cyber defenders, then expands through Google’s Fairwind Program, with no date set for developers, businesses, or consumers.
What Google actually announced
Google DeepMind unveiled Gemini 4 Carbon on October 28, 2026, just two weeks after Gemini 4 Argon, the model that opened the Gemini 4 generation. In a blog post by Koray Kavukcuoglu, SVP of Google DeepMind and Google’s chief AI architect, Google described Carbon as “our strongest model with state-of-the-art cybersecurity and long horizon capabilities.” Google CEO Sundar Pichai introduced it on X as “our next frontier model.”
Carbon is built for hard, multi-step engineering work that runs for days rather than minutes: tracking down subtle bugs across large codebases, working out root causes, and shipping fixes. That framing matters — Carbon is presented as a defender’s tool, not a general chatbot upgrade.
What is confirmed, and what is still unknown
Confirmed: Carbon was announced on October 28, 2026; it leads the Artificial Analysis Intelligence Index v5.0 at 59; its first users are a subset of cyber defenders, with wider access phased through the Fairwind Program; and its introductory API price matches Argon’s at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period.
Still unknown: Google has given no date for general availability, no context-window size, no parameter count, and no architecture details. The security scores below are Google’s own figures, measured without guardrails, and have not yet been independently reproduced.
Carbon vs. the field: the numbers that matter
The headline number is the Intelligence Index score of 59, one point ahead of Claude Opus 5.5 at 58 and two ahead of GPT-6 Astra at 57. A one-point lead is thin — but the security scores are where Carbon separates itself:
| Benchmark | Gemini 4 Carbon | Nearest rival | What it measures |
|---|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | 59 (#1) | Claude Opus 5.5: 58 | Composite across 9 evaluations |
| SEC-bench Pro | 88.2% | GPT-6 Astra: 85.4% | Finding real vulnerabilities |
| ExploitGym | 54.8% | Not published for rivals | Turning a flaw into a working exploit |
| CWE-bench | 71.0% | Opus 5.5: ~68% (reported) | Patching to root cause |
| WebDev Arena | 1768 Elo (#4) | — | Web app building, up 89 Elo points |
Google also reports internal gains over Argon: more than double the success rate at generating secure code, unsupported claims cut by roughly half, and high-severity misalignment alerts in internal coding environments down by nearly 80%. Its Chrome team reported Carbon produced high-quality patches more than 10 times better than another lab’s large frontier model — though Google did not name that lab. Treat these as Google’s claims until independent testing lands.
Comparison: Carbon vs. Argon, the model it follows
Carbon builds directly on Gemini 4 Argon, announced two weeks earlier. Argon opened to the public on October 13, 2026, in the Gemini app and paid API; Carbon has no such date. Argon scores 50 on the same Intelligence Index v5.0 where Carbon scores 59 — a nine-point jump in one generation step. The price is identical at launch, so the upgrade, when it arrives, costs no more per token than the model it replaces. The real difference is the release model: Argon went public in two weeks, while Carbon is gated on capability — specifically, its vulnerability-finding and exploitation scores.
Cost: what Carbon will cost when it reaches you
When Carbon opens more broadly, it will carry Argon’s introductory price: $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After the introductory period, that doubles to $4 and $20. At Artificial Analysis’s standard pricing, Carbon costs about $7.76 per Intelligence Index task — roughly 23% less than Claude Opus 5.5 at maximum effort ($10.04). For now, though, the price is academic: there is no app, no paid plan, and no API key you can request.
Prerequisites: who gets access, and how
Access starts this week for a subset of cyber defenders, then widens to trusted partners through Google’s Fairwind Program, which already has more than 650 partners. Priority goes to governments and national cyber authorities, critical infrastructure operators, and core technology platforms. Google runs background checks, requires phishing-resistant multi-factor authentication, and limits access to internal cybersecurity, incident response, or penetration testing teams — partners may not share or resell access. Google is also taking part in the U.S. government’s voluntary process for pre-release model testing.
Risks and limits: why Google is holding it back
The gating is the story. A model that finds real vulnerabilities at 88.2% and can turn them into working exploits is a dual-use tool: in defenders’ hands it patches systems faster, in attackers’ hands it finds the holes first. Two caveats deserve attention. First, the security benchmarks were run without guardrails, so the numbers show raw capability, not the behaviour of the shipped, safeguarded product. Second, Carbon does not lead everywhere: on the Coding Agent Index it scores 59, behind Claude Code running Opus 5.5 at 62, and on the Vals Index it sits second, just behind Sonnet 5.5. The “strongest model” claim is a security-and-long-horizon claim, not an across-the-board one.
Why it matters, and what to watch
Carbon signals a shift in how frontier models are released: the most capable models may increasingly go to vetted professionals first, with the public waiting behind them. Watch for three things: independent reproductions of the security scores, the first Fairwind partners reporting real-world results, and any date — even a quarter — for developer or consumer access.
Analysis: the gated-release precedent
Analysis, not confirmed fact: Carbon’s release looks less like a product launch and more like a controlled deployment of a dual-use capability. Google claims the benchmark crown while keeping the sharpest tool out of general circulation, buying time to learn how it behaves in defenders’ hands. Expect other labs to copy this pattern, which could split the AI market into a public tier and a vetted tier — and the open question is whether “vetted” stays a technical standard or becomes a political one.
Frequently asked questions
What is Google Gemini 4 Carbon?
Gemini 4 Carbon is Google DeepMind’s newest frontier model, announced on October 28, 2026. Google describes it as its strongest model for cybersecurity and long-horizon engineering tasks, such as finding software vulnerabilities, proving they can be exploited, and patching them across large codebases.
Can I use Gemini 4 Carbon now?
No. Carbon is not in the Gemini app, not available through a public API, and not sold to businesses or consumers. It is going first to a subset of cyber defenders, then to trusted partners through Google’s Fairwind Program, with general availability promised “as soon as possible” but no date given.
How does Gemini 4 Carbon compare to Claude Opus 5.5 and GPT-6 Astra?
Carbon leads the Artificial Analysis Intelligence Index v5.0 at 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. On security benchmarks it scores 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench. It does not lead everywhere: it trails Opus 5.5 on the Coding Agent Index and sits second on the Vals Index.
How much will Gemini 4 Carbon cost?
When it opens more broadly, Carbon will cost $2 per million input tokens and $10 per million output tokens as an introductory rate, with cached input 95% cheaper. After the introductory period the price rises to $4 and $20 per million tokens, matching Gemini 4 Argon’s pricing.
Why is Google limiting access to Gemini 4 Carbon?
Google says Carbon’s vulnerability-finding and exploitation capabilities need careful handling, so it is rolling the model out first to vetted cyber defenders while it gathers feedback and strengthens safeguards. Google is also participating in the U.S. government’s voluntary pre-release model testing process before any wider release.
Sources
- WebProNews — Google Gives Cyber Defenders First Access to Gemini 4 Carbon (reporting Google DeepMind’s October 28, 2026 announcement)
- Gulf News — Google unveils Gemini 4 Carbon with top AI benchmark score (October 29, 2026)
- DEV — Gemini 4 Carbon: Benchmarks, Pricing and What It Means for Developers
