Google releases Gemini 4 Carbon, tops the Artificial Analysis leaderboard with 59 points

According to officially released data, Gemini 4 Carbon scored 59 on the latest Artificial Analysis Intelligence Index v5.0 leaderboard, surpassing Claude Opus 5.5 and GPT-6 Astra to take the top spot among global large models.

Even more striking is that Gemini 4 Carbon can already work on complex engineering tasks for several consecutive days, autonomously finding, verifying, and fixing software security vulnerabilities, setting new records on multiple cybersecurity benchmarks.

However, ordinary users cannot yet access such a powerful model: Google has decided to give priority access to a small number of cybersecurity defenders, and there is still no clear timetable for when other users will be able to try it.

Returning to world No. 1, Gemini is also starting to push the ceiling of capability

Looking back at this year's large-model competition, Google's update cadence is clearly showing a sense of "learning from shame and then striving harder."

Previously, after successively releasing the Gemini Flash 3.5 to 3.8 series, Google was slow to deliver on expectations for a new-generation flagship Pro model. Meanwhile, Anthropic and OpenAI kept raising the capability ceiling, and domestic Chinese models were also catching up, so Google urgently needed a substantial product to prove its strength again.

It was not until two weeks ago that Gemini 4 Argon finally launched.

Argon means "argon," a stable noble gas commonly used to provide a protective environment in welding and semiconductor manufacturing. Argon achieved notable improvements in writing style and expression quality, but still fell short of fully leading the top models.

Now Google has brought out Carbon, meaning "carbon." As an important element that makes up life and organic compounds, carbon can form both soft graphite and hard diamond. (PS: At this update pace, the periodic table is no different from a product roadmap.)

What really drew attention to Gemini 4 Carbon was the series of results it announced.

On October 27, the independent model evaluation organization Artificial Analysis had just upgraded its Intelligence Index v5.0. This composite index covers four major domains—knowledge work, coding, science, and general reasoning—each weighted at 25%, with 60% of the evaluation weight using private test data or hidden answers, further increasing the difficulty of the tests.

When the new leaderboard was released yesterday, Claude Opus 5.5 ranked first with 58 points, GPT-6 Astra followed closely with 57 points, and GPT-6.1 Sol ranked third with 55 points. Gemini 4 Carbon directly scored 59 points, 9 points higher than Argon and 1 point ahead of the previous leader Claude Opus 5.5.

Beyond Artificial Analysis, the latest leaderboard released by Arena.ai also shows that Gemini 4 Carbon (High) ranks first in Text Arena, second in Agent Arena, and fourth in Code Arena WebDev.

Along with the new model topping the rankings, Google internally has also begun actively building momentum for Gemini. Chinese researcher Yao Shunyu, who previously jumped from Anthropic to Google DeepMind, also voiced support for promoting the Gemini series.

In addition to the overall ranking, Gemini 4 Carbon's performance on actual engineering tasks has also changed markedly.

In the past, large models could write a program in a few minutes, but when facing large-scale software engineering, the problem becomes much more complex. A fault hidden across hundreds of files may require developers to repeatedly check dependencies, locate the anomaly, modify the code, and then go through multiple rounds of testing before it can be resolved.

Google says Gemini 4 Carbon can already run continuously for several days, handling complex engineering problems in large codebases and low-level systems. Even when it encounters obstacles, it can keep looking for alternative approaches and ultimately locate and fix some hidden software defects.

In web development, Gemini 4 Carbon's Elo score on WebDev Arena increased by 89 points to 1,769, ranking fourth.

Compared with the programming assistants of the past that were available on demand, Google hopes Gemini 4 Carbon can become an autonomous agent that truly takes on complex engineering tasks.

However, although the previous generation Gemini 4 Argon improved in capability, its quota was consumed relatively quickly in actual use—handling just a few complex tasks could hit usage limits. It is hoped that this generation, Carbon, while raising the capability ceiling, can also improve quota consumption and usage-limit issues.

If long-horizon coding demonstrates Gemini 4 Carbon's engineering capability, then cybersecurity may be the part of Google's update most worth watching.

Traditional software security work requires going through multiple stages: first, finding vulnerabilities in a large amount of code, then determining whether a vulnerability can actually be exploited, and finally writing a fix for the problem and verifying whether the change is effective.

Gemini 4 Carbon's capability improvements happen to cover this entire workflow.

On the SEC-bench Pro test, Gemini 4 Carbon achieved a score of 88.2%, which mainly assesses the model's ability to autonomously discover complex vulnerabilities in real software;

On the ExploitGym test, Gemini 4 Carbon reached 54.8%, with a focus on whether it can actually exploit vulnerabilities to verify risk;

On Collinear's CWE-bench test, Gemini 4 Carbon achieved 71.0%, showing it can generate effective code fixes for the problems it finds.

According to results disclosed by Google, Gemini 4 Carbon reached leading levels on all three of the above tests, and the relevant scores were measured without safety guardrail restrictions enabled.

Google's internal tests also produced similar results.

In secure code generation, Gemini 4 Carbon's success rate is more than twice that of Argon; the Chrome team found that its performance in generating high-quality security patches was more than ten times higher than that of another lab's frontier model, though the comparison target was not disclosed. Google's cloud security platform Wiz also observed that Carbon achieved leading results in web application penetration testing.

From discovering vulnerabilities and verifying risk to completing fixes, AI is gradually mastering the entire workflow of cybersecurity experts. But the same capabilities can be used both to fix systems and potentially be exploited by attackers, so Google also focused on improving Carbon's reliability when autonomously executing tasks.

On the Terminal-Bench 4.0 test, Carbon's rate of successfully cheating dropped to 0.0%, and the rate of attempting to cheat was only 0.4%, while task accuracy improved.

In real programming environments, high-severity behavioral inconsistency alerts decreased by nearly 80%, assertions lacking supporting evidence decreased by about half, and unauthorized actions and cases of insufficiently completed tasks also fell by 26% and 29%, respectively.

Another advance comes from prompt injection protection.

Attackers may embed malicious instructions in web pages or documents to induce AI to deviate from its original task, and in Gray Swan's indirect prompt injection test, Carbon further reduced the attack success rate from Argon's 0.7% to 0.0%, resisting all attack attempts covered by this test.

Of course, a zero success rate in testing does not mean that all attacks in the real world can be stopped. Some Hacker News commenters believe that reducing AI's opportunistic shortcuts when executing tasks may be one of the important advances in improving the practicality of frontier models.

However, after Gemini 4 Carbon's release, the most attention-grabbing controversy was, instead, not much related to the model's capabilities.

According to the official arrangement, Gemini 4 Carbon will first be provided to some cybersecurity defenders this week, and then gradually expanded to trusted partners through the Fairwind Program. Google says it is prudently participating in the U.S. government's voluntary pre-release testing of models, and there is currently no timetable for when ordinary developers and consumers will be able to use it.

On pricing, Gemini 4 Carbon will follow Argon's launch pricing: $2 per million input tokens, $10 for output tokens, and a 95% discount on cached input. After the promotional period ends, standard pricing will be adjusted to $4 for input and $20 for output.

It is worth mentioning that when the news of Google's new model release reached Hacker News, the comment section quickly became lively.

Many commenters said outright that they are increasingly averse to Google's practice of announcing new models early while delaying access; others said they no longer even bother clicking on Google's model release links, because in the end they always find that the product cannot be used for the time being.

A flagship model that originally could have been served directly once it was polished now has to be wrapped in a safety narrative shell about being too capable and feared to destroy the world, with press releases issued every few days to hunt for everyone's attention, yet in the end it remains unavailable for a long time.

This kind of "performance without the actual product" release style has truly been led astray by Anthropic.

After all, benchmarks can tell us how impressive a model is in exams, but whether it can enter real work environments and create sustained value can only be known after users actually get to use it.

And for Gemini, perhaps what matters more than once again announcing that it is globally No. 1 is quickly giving more people the opportunity to actually use this global No. 1.