Google's newest frontier model is built to defend, not just to chat — and for now, only vetted cyber defenders get to use it.
October 29, 2026 — Google DeepMind has announced Gemini 4 Carbon, a new frontier model the company describes as its strongest yet for cybersecurity and long-running engineering work. Google announced the model on October 28, 2026, two weeks after the launch of its predecessor, Gemini 4 Argon.
Image: A glowing faceted crystal core ringed by circuit arcs and a shield outline. Credit: AI-generated illustration / Red Eye Eagle News.
Editor's Highlight: The striking part is not the benchmark scores. It is the gate: Google is putting its most capable cyber model in defenders' hands first, with no date yet for wider access — a sign of how dual-use this technology has become.
Table of Contents
- What Gemini 4 Carbon Is
- The Benchmarks Google Is Citing
- Restricted Access: Defenders First
- Pricing
- Long-Horizon Work and Secure Code
- Safety Claims and the Guardrails Caveat
- Bottom Line
- FAQ
- Sources
Key Facts
| Detail | Information |
|---|---|
| Announced | October 28, 2026, by Google DeepMind |
| Model | Gemini 4 Carbon, building on Gemini 4 Argon |
| Positioning | Google's strongest model for cybersecurity and long-horizon tasks (company claim) |
| Index score | 59 on Artificial Analysis Intelligence Index v5.0, ranked first (Google's comparison) |
| Security benchmarks | 88.2% SEC-bench Pro; 54.8% ExploitGym; 71.0% CWE-bench (Collinear) |
| Access | Subset of cyber defenders first, then trusted partners via the Fairwind Program |
| Introductory price | $2 per million input tokens; $10 per million output tokens |
| Public availability | No date given for developers, businesses or consumers |
What Gemini 4 Carbon Is
Google DeepMind calls Carbon "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." The announcement, published on the Google DeepMind blog on October 28, 2026, presents Carbon as a successor to Gemini 4 Argon, which Google says was launched two weeks earlier.
Carbon is aimed at defensive cybersecurity work: finding real vulnerabilities, validating risk by exploiting them in controlled tests, and patching flaws at their root cause. Google also presents it as a model for long, multi-step engineering tasks that can run for days on a single job.
Newsroom Focus: Every benchmark and capability figure in this article comes from Google's own announcement or from reporting of Google's comparison table. No independent test of Carbon was found in the sources read. Treat the scores below as Google's claims, not independently verified results.
The Benchmarks Google Is Citing
According to Google, Carbon ranks first on the updated v5.0 Artificial Analysis Intelligence Index with a score of 59. Google's comparison puts Anthropic's Claude Opus 5.5 at 58 and OpenAI's GPT-6 Astra at 57.
On cybersecurity tests, Google reports three headline scores. Carbon scored 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for validating risk through exploitation, and 71.0% on CWE-bench (Collinear) for patching at the root cause.
Google states that those external cybersecurity benchmarks were evaluated without guardrails. That caveat matters: the scores describe the model's raw capability on the tests, not its performance with safety restrictions in place. The caveat stays attached to every benchmark claim in this article.
For balance, Google's own comparison table shows rivals leading on some tests. GPT-6 Astra leads on FrontierSWE, at 65.5% against Carbon's 61.2%. Claude Opus 5.5 leads on OSWorld-2.1, at 87.2% against Carbon's 81.5%.
| Test | Carbon | Leader |
|---|---|---|
| Artificial Analysis Index v5.0 | 59 (1st) | Carbon |
| SEC-bench Pro | 88.2% | Carbon (Google claim) |
| FrontierSWE | 61.2% | GPT-6 Astra, 65.5% |
| OSWorld-2.1 | 81.5% | Claude Opus 5.5, 87.2% |
Restricted Access: Defenders First
Carbon is not publicly available. Google says it is engaged in the U.S. government's voluntary pre-release model-testing process, and that starting in the week of October 28, Carbon goes first to a subset of cyber defenders.
Access then expands gradually to additional trusted partners through Google's Fairwind Program. Reporting of the launch describes the priority groups as governments, critical-infrastructure operators and core technology platforms.
No date was given for developers, businesses or consumers. Google did not disclose a context window, a parameter count or an output limit for Carbon in the sources read. Those specifications for Argon must not be transferred to Carbon.
Analysis Note: Staged, defender-first access is the clearest signal of how Google views the risk. A model this good at finding and exploiting vulnerabilities is useful to attackers as well as defenders, so Google is choosing who gets it first — and saying so openly.
Pricing
Introductory pricing is $2 per million input tokens and $10 per million output tokens, the same introductory price as Argon, according to Google. Cached input tokens are priced at 95% off the input price.
After the introductory period, pricing rises to $4 per million input tokens and $20 per million output tokens, according to reporting of the launch by MarkTechPost and Daily AI Brief. Google's own blog post, in the text read, states only the introductory price, so the later figures are attributed to that reporting.
Pricing alone will not decide Carbon's reach. Access restrictions will. Until Google opens the model beyond defenders and trusted partners, the price list is a signal of intent, not an offer most readers can act on.
Long-Horizon Work and Secure Code
According to Google, internal teams have seen Carbon run for days on single multi-step engineering tasks. The examples Google gives include root-causing and patching subtle bugs across large codebases and low-level systems.
On secure code, Google says Carbon achieves more than twice the success rate of Argon at generating secure code on internal benchmarks. Google's Chrome team reported that Carbon performed more than 10 times better than another lab's large frontier model at generating high-quality patches. Google did not name the rival model, and this article does not guess at it.
Those are internal results, reported by Google. They describe what Google's own teams observed, not what independent users have reproduced. The distinction matters for a model most readers cannot yet test themselves.
Safety Claims and the Guardrails Caveat
Google reports a set of internal safety and alignment figures for Carbon. Successful cheating on Terminal-Bench 4.0 fell to 0.0%, with attempts at 0.4%. High-severity misalignment alerts in live internal coding fell by almost 80%.
Google also says unsupported claims were cut by about half versus Argon, overreach was down 26% and under-service down 29%. On Gray Swan's Indirect Prompt Injection benchmark, Carbon recorded 0.0% attack success, against Argon's 0.7%.
Practical Takeaway: Read the safety figures together with the "without guardrails" caveat on the security benchmarks. Google is reporting strong internal safety behaviour and strong raw security capability in separate tests. Neither figure, on its own, settles how Carbon will behave in the open.
Bottom Line
Gemini 4 Carbon is Google's bid to lead on cyber defence, with benchmark scores it publishes itself and access it is rationing to defenders first. The numbers are impressive as claims; the rollout is the fact. Watch who gets access next through the Fairwind Program, and whether independent testers confirm the security scores once they can run the model themselves.
FAQ
What is Gemini 4 Carbon?
It is Google DeepMind's newest frontier model, announced October 28, 2026, which Google describes as its strongest model for cybersecurity and long-horizon engineering tasks. It builds on Gemini 4 Argon, launched two weeks earlier according to Google.
Can I use Gemini 4 Carbon today?
Almost certainly not. Access goes first to a subset of cyber defenders, then gradually to trusted partners through Google's Fairwind Program. No date was given for developers, businesses or consumers.
What benchmark scores is Google claiming?
According to Google: 59 on the Artificial Analysis Intelligence Index v5.0 (ranked first in Google's comparison), 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench (Collinear). Google states the security benchmarks were evaluated without guardrails.
How much does Carbon cost?
Introductory pricing is $2 per million input tokens and $10 per million output tokens, the same as Argon, with cached input at 95% off. Reporting of the launch puts the post-introductory price at $4 and $20 per million tokens respectively.
Does Carbon beat every rival on every test?
No. In Google's own published comparison, GPT-6 Astra leads FrontierSWE and Claude Opus 5.5 leads OSWorld-2.1. Carbon's first-place claim is for the overall Artificial Analysis index, not for every individual test.
Sources
- Google DeepMind — Introducing Gemini 4 Carbon: announcement, benchmarks, guardrails caveat, rollout and pricing
- Reuters — Google launches Gemini 4 Carbon AI model with cybersecurity focus: independent confirmation of the launch, timing and rollout
- MarkTechPost — Gemini 4 Carbon benchmark table, rival-leading tests, post-introductory pricing and undisclosed specifications
- Daily AI Brief — Corroboration of the index score, benchmark scores, Fairwind rollout and pricing
- Image: AI-generated illustration created for this article / Red Eye Eagle News.
Join the conversation
Comments
Post a Comment