Gemini 4 Carbon Tops AI Index at 59, but Defenders Only

Gemini 4 Carbon is Google’s new frontier model, announced October 28, 2026. It ranks first on the Artificial Analysis Intelligence Index v5.0 with 59 points, one ahead of Claude Opus 5.5, and posts 88.2% on SEC-bench Pro. The catch: only vetted cyber defenders can use it today. Introductory API pricing is $2 input and $10 output per million tokens, doubling to $4 and $20 later.

Key takeaways

• Carbon scores 59 on the Intelligence Index v5.0, versus 58 for Opus 5.5.
• Introductory price is $2/$10 per million tokens, half of Opus 5.5’s $4/$20.
• Google says the Fairwind access program spans more than 650 partners worldwide.

What is Gemini 4 Carbon?

Gemini 4 Carbon is Google DeepMind’s strongest model, built on top of Gemini 4 Argon and tuned for cybersecurity and long, multi-step tasks. Google is releasing it in stages: cyber defenders first, then trusted partners, then developers and consumers at an unannounced date.

“Gemini 4 Carbon is our strongest model with state-of-the-art cybersecurity and long horizon capabilities,” wrote Koray Kavukcuoglu, SVP of Google DeepMind and Google’s Chief AI Architect, in the official announcement.

It is the second Gemini 4 model this month. We covered the first at launch: Gemini 4 Argon led 12 of 18 benchmarks at $2 a million. Carbon keeps Argon’s introductory price and raises the ceiling.

Internal teams saw Carbon run for days on multi-step engineering tasks, according to Google. That is the “long horizon” part of the pitch, and it matters for agent workloads billed by the token.

How good is Gemini 4 Carbon on benchmarks?

On paper, Carbon is the best general model available to anyone right now, by a single point. Artificial Analysis lists Claude Opus 5.5 at 58 and GPT-6 Astra at 57 on Intelligence Index v5.0. Google says Carbon scores 59. Its security numbers are the bigger leap.

Security scores

Google reported 88.2% on SEC-bench Pro, which tests autonomous discovery of complex flaws. It reported 54.8% on ExploitGym, which checks whether a found bug can actually be exploited, and 71.0% on Collinear’s CWE-bench for root-cause patching.

Google’s Chrome team says Carbon was “more than 10X better than another lab’s large frontier model” at generating high-quality patches. Google did not name the lab. Internally, it claims more than 2x the secure-code success rate of Argon.

Coding and general scores

On WebDev Arena, Carbon gained 89 Elo points to reach 1768, which puts it fourth, not first. That gap is worth noting. A model that leads the composite index still trails three rivals in head-to-head web app building, as voted by users.

Model	Intelligence Index v5.0	Input $/M tokens	Output $/M tokens	Who can use it
Gemini 4 Carbon (intro)	59	$2	$10	Vetted cyber defenders
Gemini 4 Carbon (standard)	59	$4	$20	Not yet available
Claude Opus 5.5	58	$4	$20	General API
GPT-6 Astra	57	$10	$50	General API

How much does Gemini 4 Carbon cost?

Gemini 4 Carbon costs $2 per million input tokens and $10 per million output tokens during an introductory period, with cached input 95% off. After that, Google’s footnote lists $4 input and $20 output. Google has not said when the introductory window ends.

The standard price is identical to Claude Opus 5.5, which Artificial Analysis lists at $4 and $20. So the real offer is simple: during the intro window, Carbon is half the price of its nearest rival for a model that scores one point higher.

Against GPT-6 Astra, the gap is wider. Astra lists at $10 input and $50 output, according to Artificial Analysis, five times Carbon’s intro rate. OpenAI has been cutting prices lower in the stack, as we noted when GPT-6.1 Terra halved its price to $1 per million, but its top model remains expensive.

Here is the skeptical part. A price nobody outside a vetted program can pay is a marketing number. By the time ordinary developers get access, the introductory window may be gone, and Carbon would simply match Opus 5.5 on price.

Who can use Gemini 4 Carbon right now?

Only a subset of cyber defenders, starting this week. Access widens gradually through Google DeepMind’s Fairwind Program. Developers, businesses and consumers come later, “as soon as possible,” with no date. Google also says it is taking part in the U.S. government’s voluntary pre-release model testing.

Fairwind is not an open beta. According to Unite.AI’s summary of the program terms, participants accept strict conditions:

• Only permitted dual-use work, such as authorized threat simulation, reverse engineering and defensive malware analysis.
• No malware creation, and no sharing, redistribution or resale of access.
• User-level authentication with phishing-resistant multi-factor login.
• Access limited to security teams, with employee access tracking and background checks.
• Priority for governments, national cyber authorities, critical infrastructure and core technology platforms.

The staged rollout mirrors a pattern across the industry. When OpenAI’s model crossed its own “critical” cyber line in September, the debate shifted from whether such models ship to who gets them first.

Is Gemini 4 Carbon safer than Argon?

Google says yes, on every metric it published. Carbon posted a 0.0% attack success rate on the Gray Swan Indirect Prompt Injection Benchmark, versus 0.7% for Argon. Successful cheating on Terminal-Bench 4.0 fell to 0.0%, with attempts down to 0.4%.

Google also reports roughly 2x fewer unsupported claims than Argon, and high-severity misalignment alerts in live internal coding down almost 80%. In multi-step plans, overreach fell 26% and under-service fell 29%.

Prompt injection is the number enterprise buyers should care about most. Agents that read email, tickets and web pages are the attack surface behind incidents like the Anthropic breach where AI agents got in under 12 hours. A true 0.0% would be a selling point. It is also Google’s own count on one benchmark.

Who wins and who loses from Gemini 4 Carbon?

Google wins mindshare and enterprise security budgets. Anthropic and OpenAI lose the “best model” headline and face a price fight at the top. Security vendors face a model that finds, exploits and patches bugs, though only for customers Google approves.

Winners

Alphabet. Google needs Cloud and API revenue to offset a slowing core. Alphabet reported $127.6 billion in quarterly revenue, up 25%, while Search ad growth slowed from 19% in Q1 to 12% in Q3, according to The Neuron’s daily digest. Spending is the other side: we covered how Alphabet free cash flow hit -$9.1 billion as Cloud doubled. A top-ranked model helps justify that bill.

Wiz. Google says Wiz reports Carbon leads its web penetration testing benchmarks. Google agreed to buy Wiz for $32 billion in 2025, so a Carbon-powered security stack is an obvious upsell path. Wiz published no figures.

Losers

Anthropic. Opus 5.5 lost first place by one point and now faces a rival at half its price during the intro window. Anthropic is heading toward a public listing, and pricing power at the top tier is a core part of that story.

OpenAI. GPT-6 Astra sits third on the index at five times Carbon’s intro price. OpenAI can defend share with discounts, but each cut compresses margins on its most expensive model.

Point-solution security vendors. Code scanners and pen-testing shops sell exactly what Carbon claims to automate. The threat is real but slow: Fairwind gates access, and buyers will want independent proof first.

Should you trust Google’s Gemini 4 Carbon numbers?

Treat them as claims, not results. Nearly every security and safety figure comes from Google. Outside researchers cannot test Carbon yet, and the cyber evaluations were run without guardrails, so deployed behavior may differ. None of the headline numbers have been reproduced outside Google, as AI Weekly noted.

Three specific gaps stand out. The “10X better” patch claim compares Carbon with an unnamed model. The Wiz claim has no numbers. And the one-point index lead sits inside the range where a model update from a rival flips the ranking.

What to watch next

Google has given no dated milestones for Carbon beyond this week’s first access. So the signals matter more than the calendar. These are the ones that will tell you whether Carbon changes AI pricing or stays a security niche product.

1. A general API date. Watch the Gemini API and Vertex AI model lists for Carbon outside Fairwind.
2. The end of the intro window. If developers only ever see $4/$20, the price advantage over Opus 5.5 disappears.
3. Independent cyber tests. Third-party SEC-bench or ExploitGym runs, or a government testing summary, would validate or puncture the claims.
4. Rival responses. A price cut or new top model from Anthropic or OpenAI would reset the one-point lead.

FAQ

When was Gemini 4 Carbon released?

Google announced it on October 28, 2026. Access started that week for a subset of cyber defenders.

Can I use Gemini 4 Carbon in the Gemini app or API?

Not yet. Google says it will reach developers, businesses and consumers “as soon as possible” but gave no date.

How much does the Gemini 4 Carbon API cost?

$2 per million input tokens and $10 per million output tokens introductory, with cached input 95% off. The later standard rate is $4 and $20.

Is Gemini 4 Carbon better than Claude Opus 5.5?

By one point on the Intelligence Index v5.0: 59 versus 58. Carbon ranks fourth on WebDev Arena, so it does not lead everywhere.

What is the Fairwind Program?

Google DeepMind’s vetted access program for security teams, which Google says includes more than 650 partners. It prioritizes governments and critical infrastructure.

Why is Google restricting access?

Carbon can find and exploit software flaws. Google wants defenders to get it before attackers, and is iterating on guardrails first.

The bottom line

Gemini 4 Carbon puts Google back on top of the leaderboard, but by one point and on its own testing. The money story is price: at $2/$10, Carbon undercuts Opus 5.5 by half. That only matters once developers can buy it, and Google has not said when.

Our verdict: a strong launch for defenders, a pricing threat to Anthropic and OpenAI, and not yet a product for everyone else. Judge it when the general API ships and the intro price is still on the table.

Sources

• Google: Gemini 4 Carbon, our next frontier model
• Artificial Analysis: Claude Opus 5.5 vs GPT-6 Astra
• Unite.AI: Staged rollout sends Gemini 4 Carbon to cyber defenders first
• AI Weekly: Google ships Gemini 4 Carbon, tops AA Intelligence Index at 59
• The Neuron: Everything that happened in AI today, October 29, 2026
• Free Malaysia Today / Bernama: Google unveils Gemini 4 Carbon AI model

Wealth Engine researches and drafts with AI tools and checks every figure against the sources above. How we report.