What Is Gemini 4 Carbon? Google's Frontier Model for Agents and Security (2026)

What Is Gemini 4 Carbon? Google's Frontier Model for Agents and Security (2026)

TL;DR

• Gemini 4 Carbon is Google DeepMind's newest frontier model, announced on October 28, 2026. Google calls it its "strongest model with state-of-the-art cybersecurity and long horizon capabilities" (Google blog).
• It ranks #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59, per Google and independent coverage (Google blog, VentureBeat).
• Headline security results: 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench (Google blog).
• It builds on Gemini 4 Argon, which launched about two weeks earlier.
• You probably can't use it yet. Access starts with a subset of cyber defenders and expands through Google's Fairwind Program. Broader availability for developers, businesses, and consumers comes "as soon as possible," with no date given.
• Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% off. VentureBeat reports later prices of $4 and $20 (VentureBeat).

What is Gemini 4 Carbon?

Gemini 4 Carbon is the latest model in Google's Gemini 4 family. Koray Kavukcuoglu, SVP at Google DeepMind and Google's Chief AI Architect, introduced it in the official announcement as "the next frontier model."

Two things set it apart in Google's framing:

1. Long-horizon work. Carbon is meant for long, multi-step tasks, especially in software engineering and agentic workflows.
2. Cybersecurity defense. Google claims state-of-the-art results across the defense lifecycle: finding vulnerabilities, checking whether they can be exploited, and patching them.

Carbon follows Gemini 4 Argon, which Google opened up more broadly on October 13 (VentureBeat). Google presents the two-week gap as evidence of "the pace at which we're pushing the frontier forward."

If you've followed this season's other launches, like OpenAI's split release covered in our GPT-6.1 Sol vs Terra explainer, the pattern will look familiar. Frontier labs now ship faster and use staged access for their most capable models.

Key capabilities

Long-horizon agents and coding

Google says Carbon shows "significant gains across coding and agentic tasks" (Google blog). According to The Verge, Google says its internal teams have seen the model "run for days to persist through dead ends to tackle complex, multi-step engineering tasks."

Google also says Carbon is better at:

• avoiding shortcuts when it solves tasks,
• resisting prompt injection attacks, which matters a lot for agents that read untrusted web pages, emails, or files,
• collaborating with users instead of charging ahead.

On honesty, The Verge reports Google's claim that Carbon cuts "unsupported claims" by 2x compared with Argon. Google's blog adds that in live coding environments with its own teams, Carbon shows "reduced misalignment and fabrication rates."

Cybersecurity defense

Security is the focus of this launch. Google's announcement lists three stages where it says Carbon leads (Google blog):

• Finding vulnerabilities: 88.2% on SEC-bench Pro, described as autonomously uncovering complex flaws in real-world software.
• Validating risk through exploitation: 54.8% on ExploitGym, which helps defenders decide which bugs to fix first.
• Patching root causes: 71.0% on Collinear's CWE-bench, turning findings into working code fixes.

Google also cites some internal and partner results:

• More than 2x uplift in success rate for generating secure code over Argon on internal benchmarks.
• The Chrome team reported that Carbon performed more than 10x better than "another lab's large frontier model" at generating high-quality patches. Google doesn't name the other model.
• Wiz, which Google calls its leading cloud security platform, sees Carbon leading on web penetration testing benchmarks.

Google notes that these external cyber benchmarks were evaluated without guardrails. The model you eventually get through a product may behave differently.

Benchmarks (verified)

The table below lists selected rows from the comparison table on Google DeepMind's Gemini page. All figures are as published by Google. The AA Index row comes from Artificial Analysis.

What stands out:

• Carbon tops most rows, but not all of them. VentureBeat counts the highest result in 15 of 17 rows in Google's table, with OpenAI ahead on one software engineering test and Anthropic ahead on one computer-use test (VentureBeat). The table above shows this: GPT-6 Astra leads on FrontierSWE 2 and Claude Opus 5.5 leads on OSWorld-2.1.
• The biggest jump over Argon is in long-context reasoning (GraphWalks) and exploitation (ExploitGym).
• The AA Index margin is small. Carbon's 59 is one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra.

Google publishes its methodology on DeepMind's evals methodology page for Gemini 4 Carbon, which is linked from the model page.

Availability: how to access Gemini 4 Carbon

This is the part that matters most if you want to try it. As of launch:

• Not generally available. Google says it is "taking a deliberate approach to rollout" and is taking part in the U.S. government's voluntary process for pre-release model testing (Google blog).
• First access: cyber defenders. "Starting this week," Carbon goes to a subset of cyber defenders.
• Then: the Fairwind Program. Google will "gradually phase access to additional trusted partners through our Fairwind Program."
• Later: everyone else. Google says it will keep iterating on guardrails before making Carbon available to "developers, businesses, and consumers via our product surfaces as soon as possible." It hasn't given a date.

Pricing: Carbon will launch at the same introductory price as Argon: $2 per million input tokens, $10 per million output tokens, and cached input 95% off the input price (Google blog). VentureBeat reports a footnote listing subsequent prices of $4 and $20, with no stated end date for the introductory period (VentureBeat).

In the meantime, Gemini 4 Argon is the model most teams can actually evaluate.

Gemini 4 Carbon vs Argon

Google says Carbon delivers more than a 2x uplift over Argon in secure code generation and cuts unsupported claims in half (Google blog, The Verge).

How it compares to other frontier models

Based on the published numbers:

• vs Claude Opus 5.5: Carbon leads by one point on the AA Index and on most agentic and cyber rows, but Opus 5.5 is clearly ahead on OSWorld-2.1 computer use (87.2% vs 81.5%).
• vs GPT-6 Astra: Carbon leads on the AA Index, Terminal-Bench, and all three cyber benchmarks. Astra leads on FrontierSWE 2 (65.5% vs 61.2%).
• Professional tasks: VentureBeat reports that independent testing shows Carbon trailing several rivals on professional tasks, even though it beats Argon on coding and knowledge work (VentureBeat).

Bottom line: on paper Carbon is the strongest all-rounder right now, but the gaps are narrow, and the "best" model still depends on the task.

Limitations and open questions

• Limited access. Most readers can't run it yet, so independent hands-on testing is thin.
• Hallucination trade-off. Google claims fewer unsupported claims than Argon. Even so, VentureBeat reports that independent testing recorded a higher hallucination rate than Argon on a factual-knowledge test (VentureBeat). Treat factual outputs with care.
• Guardrails will change behavior. The cyber benchmarks were run without guardrails. The public model will ship with them, by design.
• Vendor-reported numbers. Most figures come from Google's own table. Artificial Analysis gives independent confirmation for the AA Index ranking.
• Pricing will rise. The $2/$10 rate is introductory.
• Dual-use concerns. A model that's good at finding and exploiting vulnerabilities is useful to defenders and attackers alike. That's why the rollout is staged.

FAQ

What is Gemini 4 Carbon?

It's Google DeepMind's newest frontier model, announced October 28, 2026. It focuses on cybersecurity defense and long-horizon agentic and coding tasks, and it ranks #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59.

Can I use Gemini 4 Carbon today?

Probably not. It's going first to a subset of cyber defenders and then to trusted partners through the Fairwind Program. Google says it will reach developers, businesses, and consumers "as soon as possible," with no date given.

What is the Fairwind Program?

It's the program Google is using to phase Carbon access to trusted partners after the initial group of cyber defenders. Google hasn't published detailed eligibility criteria in its announcement.

How much does Gemini 4 Carbon cost?

At launch it costs $2 per million input tokens and $10 per million output tokens, with cached input 95% off. That's the same introductory price as Argon. VentureBeat reports later prices of $4 and $20.

Gemini 4 Carbon vs Argon: what's the difference?

Carbon is the newer and stronger model: 59 vs 50 on the AA Index, better cyber and long-context scores, and more than 2x better secure code generation on Google's internal tests. Argon is the one you can broadly access today.

Conclusion

Gemini 4 Carbon is Google's strongest bid yet for frontier leadership. It leads on the updated AA Index, sets top results on several cybersecurity benchmarks, and is built for agents that run for days. It's also a sign of where frontier releases are heading: staged, security-first rollouts, with the most capable models going to vetted users before anyone else.

For most builders, that means evaluating Argon now and watching for Fairwind and API updates. If you're curious how assistants behave when you push for judgment instead of limits, see our piece on Grok Bot soul prompts vs jailbreaks. And if you're after open-ended creative chat without the strict filters mainstream assistants apply, HackAIGC is built for that.