This model has achieved state-of-the-art performance in both cybersecurity and long-range task processing.Article author and source: MachineHeart
Just now, Google announced the launch of its next-generation frontier model, Gemini 4 Carbon, dubbed the "most powerful model."

The model has achieved state-of-the-art performance in cybersecurity and long-range task processing. Carbon can handle complex tasks continuously for days without stalling or taking shortcuts to complete them, ultimately delivering high-quality, reliable code.
Only two weeks have passed since the release of the previous flagship, Gemini 4 Argon. Demis Hassabis opened with a lightning emoji, indicating that Google is "iterating at lightning speed" on its new model.

Great, very confident. According to the data, this new Carbon model is truly impressive. Carbon scored 59 points on the Intelligent Index v5.0 from the third-party evaluator Artificial Analysis, ranking first. Claude Opus 5.5 followed closely with 58 points, and GPT-6 Astra came in third with 57 points. Although the three models are only two points apart, first place is still first place.

In addition, Carbon achieved 88.2% on the vulnerability detection benchmark SEC-bench Pro, 54.8% on the exploit generation benchmark ExploitGym, and 71.0% on the vulnerability fix benchmark CWE-bench, establishing a clear lead in all three benchmarks.

Blog link: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
End-to-end cybersecurity SOTA, Chrome team tested "more than 10x better"
Google devoted significant space in its official blog to Carbon's cybersecurity capabilities. Specifically, Carbon excels in all three critical phases of cybersecurity defense, achieving SOTA performance in each.
Finding vulnerabilities: On the SEC-bench Pro benchmark, Carbon achieved a score of 88.2%, surpassing GPT-6 Astra’s 85.4% and Argon’s 83.4%. This benchmark tasks models with autonomously discovering complex security vulnerabilities hidden within real-world software.

Vulnerability validation: On ExploitGym, Carbon scored 54.8%, significantly outperforming second-place GPT-6 Astra at 42.4%. This test measures a model’s ability to verify the actual exploitability of vulnerabilities, helping security teams prioritize patching. Within a 10-hour time limit, Carbon’s score differed only slightly from its 6-hour result (54.8% vs. 53.5%), indicating that it completes the majority of validation tasks quickly.

Fixing vulnerabilities: After identifying a vulnerability, addressing its root cause is a critical step in the cybersecurity defense lifecycle. Carbon achieved 71.0% on Collinear’s CWE-bench, directly converting validated findings into actionable code fixes.

In addition to public benchmarks, Google also provided internal data: Carbon achieved more than twice the success rate of Argon in generating secure code. Feedback from the Chrome browser team was even more striking—they reported that Carbon performed over 10 times better than “a large state-of-the-art model from another lab” in generating high-quality security patches.
Google’s cloud security platform, Wiz, also participated in the testing and observed that Carbon led in the web penetration testing benchmark, one of the core capabilities of Wiz’s “Scan for Good” initiative.
Long-term task: Can run continuously for several days
Another major highlight of Carbon is its "long-range task" capability, which is also becoming one of the key competitive abilities among next-generation frontier large models.
Google states that its internal team observed Carbon running continuously for several days on large codebases and underlying systems, performing root-cause analysis and patch generation across files. When encountering dead ends, the model autonomously adjusts its strategy to continue progressing without terminating the task.
One of the technical foundations of this capability is a significant improvement in long-context processing. In the GraphWalks test (BFS F1 evaluation across 256k to 1M tokens), Carbon surged from Argon’s 84.2% to 96.2%, outperforming GPT-6 Astra’s 71.8% and Claude Opus 5.5’s 66.8% by more than 20 percentage points. With a 1M token context window, the model can process the equivalent of approximately 1,500 pages of A4 documents in a single pass.
Long-term capability is becoming increasingly important because agent workflows are placing higher demands on models. For example, in software engineering, truly complex development tasks often involve continuously tracking code changes, coordinating modifications across multiple branches, managing cross-module dependencies, and repeatedly testing and fixing issues. These tasks can last for hours or even days, far exceeding the scope of a single question-and-answer interaction.
For cutting-edge large models, the ability to maintain goal consistency, correct errors promptly, and ultimately deliver reliable results over extended execution periods is becoming a key standard for measuring their practical engineering value.
Benchmark Overview: Smart Index Evaluated Across 9 Metrics, Ranking First in 4 Categories
Artificial Analysis Smart Index v5.0 is currently one of the most widely referenced comprehensive evaluation systems in the industry.
Carbon ranked first with a total score of 59, achieving first place in four categories (HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA), surpassing Claude Opus 5.5 (58 points) by 1 point, GPT-6 Astra (57 points) by 2 points, and improving by 9 points from Argon’s score of 50 two weeks ago. Artificial Analysis noted that Carbon’s lead is driven by strong performance in complex reasoning, knowledge, and scientific tasks, with significant improvements in programming and knowledge work—particularly compared to historical Gemini series performance—though it has not yet achieved comprehensive leadership in these areas.

The following are broken down by dimension.
In programming, traditional weak points of Google’s models have been notably improved by Carbon. On the Artificial Analysis Coding Agent Index, Carbon paired with Antigravity CLI (high reasoning) scored 59, just 3 points behind the top-ranked Claude Code + Claude Opus 5.5 (max reasoning, 62). In individual benchmarks, Carbon achieved 64% on Terminal-Bench 4.0 and 51% on FrontierCode v1.1, ranking second in both, behind only Claude Opus 5.5 (max, 56%). It also rose 89 Elo points in WebDev Arena (full-stack web development arena) to claim fourth place (1768 Elo).

Knowledge work: On AA-Briefcase, Carbon (high) scores 1609 Elo, an increase of 126 points over Argon (high, 1483). The largest improvement is in the Presentation dimension, rising 293 Elo to 1598; Analytical Quality also increased by 120 points to 1671. However, it still lags behind Kimi K3.1 Pro (1678), Muse Wonder (max, 1659), and Grok 4.7 (xhigh, 1638).

Knowledge reliability and hallucination rate: AA-Omniscience v1.1 scores 54 on Carbon, outperforming Claude Opus 5.5 (max, 48) by 6 points and Argon (46) by 8 points. Accuracy (percentage of correct answers) increased from 48% for Argon to 58%, a 10-percentage-point improvement. The hallucination rate rose from 4% for Argon to 11%, representing an increase, but remains the second lowest among all models, behind only Argon. For comparison, both Claude Opus 5.5 (max) and GPT-6 Astra (max) have hallucination rates of 50%.

Science and Mathematics: RiemannBench 80.0%, Terminal-Bench-Science 64.8%, both current highest scores.
Cost efficiency: Carbon’s average cost for a smart index task is $7.76, approximately 23% lower than Claude Opus 5.5 (max, $10.04) and 1.2 times that of GPT-6 Astra (max, $6.25). It generates an average of 123k tokens per task, 29% fewer than Claude Opus 5.5 (max, 173k). Compared to Argon at the same price point ($7.68 per task), costs are nearly identical, but Carbon delivers a smart index score 9 points higher.

Artificial Analysis Model Evaluation: https://artificialanalysis.ai/models/gemini-4-carbon
Security alignment: Cheating rate 0%, prompt injection breach rate 0%
Carbon builds upon Argon’s multi-layered defense architecture for security alignment, including alignment during training, real-time abuse detection, prompt injection defense, and chain-of-thought monitoring, with further enhancements.
The cheating rate has dropped to 0.0%. In the Terminal-Bench 4.0 test, Carbon achieved a 0% success rate for cheating attempts, with only 0.4% of attempts involving cheating, while overall task accuracy continues to improve. Google states this was accomplished by further mitigating shortcut exploitation during the training phase.
Prompt injection attack success rate: 0.0%. In the Gray Swan indirect prompt injection (IPI) benchmark (k=15 attack attempts), Carbon successfully defended against all attacks. In comparison, other state-of-the-art models performed as follows: Argon at 0.7%, Claude Opus and Claude Fable at 1.0% each, GPT-6 Astra at 8.5%, and Grok 4.6 at a high 51.8%.

At the behavioral control level, Carbon has improved two common alignment failure modes: overreach (operations exceeding user intent) decreased by 26%, and under-service (omission of requested operations) decreased by 29%. Google describes Carbon’s behavioral adjustment when facing ambiguous tasks as actively retrieving relevant sources, presenting information discrepancies to the user, and only acting after confirmation—avoiding guesses or forced actions.
In terms of model transparency, Carbon makes approximately half as many unsupported claims as Argon. When critical information is missing, the model explicitly notes methodological limitations or lists incomplete prerequisites in its output, rather than fabricating or obfuscating details.
Internal team testing shows that high-severity aligned alerts have decreased by nearly 80% compared to Argon.
Pricing and Availability
Carbon is priced the same as Argon, at $4 per million input tokens and $20 per million output tokens. A 50% launch discount is currently available, with no end date announced. Cached input tokens receive a 95% discount, reducing the cost to just $0.20 per million. The context window supports up to 1M tokens and accepts text, images, video, and audio inputs.
In terms of cost per task, Carbon averages $7.76 per intelligent index task, approximately 23% lower than Claude Opus 5.5 ($10.04), but about 24% higher than GPT-6 Astra ($6.25). It generates an average of 123k tokens per task, which is 29% fewer than Claude Opus 5.5’s 173k, indicating that Carbon also has an advantage in token efficiency.
Starting this week, Carbon is initially opening to trusted network defenders in the Fairwind project, and will gradually expand access after gathering feedback and refining security safeguards, before fully opening to developers, businesses, and consumers.
Epilogue
The familiar cycle has returned. Just as Claude dropped a "nuclear bomb," Google served up a new dish, and OpenAI ramped up its efforts. After a full circle, it’s another unremarkable week in the AI industry.
Earlier, Google seemed to lose its voice amid the rapid iteration of cutting-edge large models, leading some to question whether it could still remain competitive. However, its recent consecutive breakthroughs, especially Carbon, have once again drawn attention.
Of course, as giants compete fiercely with one another, the entire industry benefits. But rather than generating buzz through doomsday narratives about "AI destroying the world" or exaggerated marketing that leaves people dizzy and overwhelmed, we hope to see tangible progress from models in real-world tasks.
Less showmanship, more real skill. Such an AI competition is truly worth looking forward to.



