Google DeepMind Releases Gemini 4 Carbon Two Weeks After Argon, Starting With Cyber Defenders

Google DeepMind announced Gemini 4 Carbon on October 28, 2026, the successor to Gemini 4 Argon, which reached the market two weeks earlier. The company calls it its next frontier model and says it is the best at cybersecurity and long-horizon tasks. Access is limited, though: the model goes first to a narrow group of defenders.

According to a Google blog post by Koray Kavukcuoglu, Carbon is meant for hard, multi-step engineering tasks. Teams inside the company ran it on jobs lasting several days, such as finding the root causes of subtle bugs and patching them in large codebases.

Cybersecurity in the spotlight

Google reports results on three benchmarks: 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploiting them) and 71.0% on CWE-bench (patching). According to The Decoder, some of these tests were run without the model's safeguards, and Carbon leads the SEC-bench Pro ranking.

The company also says that in internal tests Carbon generates secure code more than twice as often as Argon. Google's Chrome team reportedly saw patches more than ten times better than those from another lab's model. The Decoder notes that these are Google's internal claims and cannot be independently verified.

Independent benchmark results

On the Artificial Analysis Intelligence Index v5, Carbon scored 59 points and took first place, while Argon had 50. In agentic coding the model improved from 53 to 59 points, but according to The Decoder it still trails Claude Code with Opus 5.5 (62) by three points.

On Arena.ai, the model leads Text Arena with 1528 points, is second in Agent Arena behind Opus 5.5 and fourth in WebDev Arena. On Vals Index v2.2 it scored 70.36%, second behind Sonnet 5.5 (70.85%) and ahead of Opus 5.5 (70.18%).

Costs and caveats

The introductory price is the same as Argon's, and cached input tokens are 95% cheaper. Artificial Analysis calculates costs at list prices ($4 and $20) and arrives at $7.76 per task, versus $7.68 for Argon and $10.04 for Opus 5.5.

There are drawbacks too. In Vals AI tests, finance and tax tasks take more than twice as long as with Argon, and the model generates 2.3 times more reasoning tokens and 64% more output tokens than Argon, though still less than half of what Sonnet 5.5 and Opus 5.5 produce. The hallucination rate rose from 4% to 11%, although according to The Decoder it is still the second lowest in the ranking, and accuracy grew from 48% to 58%.

Today, we’re announcing the next frontier model: Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long horizon capabilities. - Google DeepMind, Google blog post

Limited access

Google is taking part in the U.S. government's voluntary process for pre-release model testing. This week Carbon goes to a subset of cyber defenders, and then to additional trusted partners through the Fairwind Program. The company says it will make the model available to developers, businesses and consumers as soon as possible, once it has improved the safeguards based on feedback from the first group.

For Polish security teams and developers, that means watching the results from a distance for now. The model is not generally available, and no specific date for a wider rollout has been given.