Google Unveils Gemini 4 Carbon: A Quantum Leap in Cybersecurity
On October 28, 2026, Google introduced Gemini 4 Carbon, touting it as its most formidable model designed for the realms of cybersecurity and long-term analytical endeavors.
However, the rollout of this innovation will be gradual. Initially, access will be granted to a select group of cyber defenders, with broader availability for developers, enterprises, and consumers to follow in due course.
This first cohort of cyber defenders remains unspecified, as Google opts for discretion regarding the collaborating entities. The expansion of access is expected to follow a phased approach via its Fairwind Program, aimed at trusted partners.
Additionally, Google is participating in the U.S. government’s voluntary pre-release testing for model assessment.
A promise is made that developers, enterprises, and individuals will gain access to Carbon through Google’s product interfaces “as soon as possible”; however, specific timelines remain unspecified.
Gemini 4 Carbon builds upon the foundation laid by the previous model, Gemini 4 Argon. “Carbon exemplifies the rapid progression we are making in advancing the frontiers of artificial intelligence,” remarked Koray Kavukcuoglu, Senior Vice President of Google DeepMind and Chief AI Architect.
Google positions Carbon as an agentic model capable of sustained operational efficiency. Internal reports indicate that Carbon can execute complex, multi-step engineering tasks over extended periods, successfully navigating various challenges.
Notable applications include identifying root causes and rectifying subtle bugs across expansive codebases and intricate systems—though these insights stem from internal observations rather than peer-reviewed performance tests.
The pricing structure for Carbon mirrors that of Argon, commencing at $2 per million input tokens and $10 per million output tokens.
An especially cost-effective option is available for cached input tokens, priced at a mere 5% of the standard input token cost.
A footnote reveals post-introductory rates: $4 per million input tokens and $20 per million output tokens, though the timeline for this transition remains undisclosed, as does any consumer subscription information.
Analyzing Google’s Benchmark Scores
In the latest iteration of the Artificial Analysis Intelligence Index (v5.0), Google positions Carbon at the pinnacle with a score of 59.
The model also performs commendably across three cybersecurity benchmarks, each evaluated under unregulated conditions.
Carbon attained an impressive 88.2% on the SEC-bench Pro, a figure self-calculated as pass@1 over a 183-task split of a May 2026 JavaScript engine, averaged across three attempts.
Furthermore, on ExploitGym, it registered a score of 54.8%, again self-computed with one attempt per task, devoid of internet connectivity and with mitigations disabled.
This success rate was determined under both standard six-hour and extended ten-hour timeframes. Additionally, Carbon achieved a 71.0% score on Collinear’s CWE-bench, corroborated by that benchmark’s public leaderboard.
The conditions allowing for unregulated evaluation applied solely to these three cybersecurity benchmarks, not universally across Google’s published comparisons.
Competitors also feature in the compilation; for instance, GPT-6 Astra scored 65.5% on FrontierSWE 2 compared to Carbon’s 61.2%.
Meanwhile, Claude Opus 5.5 registered a score of 87.2% on the OSWorld-2.1 offline subset, outpacing Carbon’s 81.5%.
The benchmark table is a Google-curated collection, incorporating various sources. Notably, Google calculated only some of these figures, which run on the Gemini API under optimal settings.
The evaluation methodology clarifies exceptions to this processing. Several rows stem from external evaluators, including the Artificial Analysis Index and CWE-bench from Collinear AI.
Other scores derive from self-reported data from providers, unless specified otherwise. These results are current as of October 2026.
Internal Evaluations of Carbon’s Performance
Numerous assertions regarding Carbon’s efficacy originate from within Google. The company claims that Carbon has dramatically more than doubled the success rate in generating secure code compared to Argon.
Moreover, the Chrome team noted that Carbon surpassed “another lab’s large frontier model” in high-quality patch generation by a factor of over ten, though specific details regarding the lab or the benchmark remain undisclosed.
Wiz, acknowledged as Google’s premier Cloud security platform, has also recognized Carbon’s superiority in benchmarks for web penetration testing.
On the safety front, Google asserts that Carbon has successfully slashed the incidence of successful exploits on Terminal-Bench 4.0 to 0.0%, with attempts plummeting to 0.4%.
During live coding scenarios with internal teams, high-severity misalignment alerts reportedly dropped by nearly 80%.
Internal evaluations indicated a reduction in unsupported claims by approximately double compared to Argon, alongside a 26% decrease in overreach and a 29% reduction in under-service, despite a lack of clarity on these specific measures.
In third-party testing, Gray Swan’s Indirect Prompt Injection benchmark recorded a 0.0% attack success rate for Carbon, compared to Argon’s 0.7%.
Notably absent from the announcement are certain crucial details that prospective users might find pertinent.
There is a lack of a general availability date, as well as no information on context window limits, output token caps, or a Carbon model card.
Additionally, the linked report on safety and security pertains to Argon, raising concerns regarding the internal metrics presented, which lack external validation.
Source link: Techmymoney.com.
