Gemini 4 Carbon takes the lead in AI benchmarks, but Google is reserving its model for cybersecurity experts

Google unveils Gemini 4 Carbon, its most advanced model, topping intelligence and cybersecurity leaderboards, with restricted initial access.

88.2% for detecting software vulnerabilities, 54.8% for exploiting them, and 71% for producing fixes. Google DeepMind reveals Gemini 4 Carbon, a new artificial intelligence model that places cybersecurity at the center of its advancements. It also takes the top spot in the Artificial Analysis Intelligence Index v5.0 leaderboard with a score of 59, ahead of Claude Opus 5.5 and GPT-6 Astra.

Announced on October 28, 2026, Carbon succeeds Gemini 4 Argon, which was introduced a few weeks prior. Google describes it as its most capable model, designed for complex tasks that require long sequences of reasoning, coding, and tool use.

However, unlike a traditional launch, Gemini 4 Carbon is not immediately accessible to everyone. Google is reserving its initial deployments for a limited group of defensive cybersecurity professionals as part of its Fairwind program. Google Reclaims Top Spot on the Intelligence Index

Gemini 4 Carbon achieves 59 points on version 5.0 of Artificial Analysis's Intelligence Index.

This score places it ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57, and Gemini 4 Argon at 50.

The ranking, established by an independent organization, measures several performance dimensions of the models, including reasoning, knowledge, mathematics, and coding.

However, the gap between the top three models remains relatively small. A single point on a composite index does not necessarily mean a noticeable difference on every real-world task.

In its comparison table, Google also presents strong results on more specialized evaluations.

Benchmark · Gemini 4 Carbon · Gemini 4 Argon · GPT-6 Astra · Claude Opus 5.5 AA Intelligence Index v5.0 · 59 · 50 · 57 · 58 Terminal-Bench 4.0 · 65.2% · 57.1% · 59.1% · 59.6% Agents' Last Exam v1.1 · 77.4% · 69.0% · 73.1% · 71.6% GraphWalks, long context · 96.2% · 84.2% · 71.8% · 66.8% RiemannBench · 80.0% · 76.0% · 72.0% · 69.6%

These results show significant progress compared to Argon, particularly for agents and processing large contexts.

However, Carbon does not dominate every test. On FrontierSWE 2, GPT-6 Astra reaches 65.5%, compared to 61.2% for Carbon. On OSWorld-2.1, Claude Opus 5.5 maintains the lead with 87.2%, compared to 81.5%.

A distinction must therefore be made between holding first place on an overall index and overall superiority across all domains. A Leap Forward in Cybersecurity

It is in cybersecurity that Google emphasizes its new model's progress the most.

The company organizes its results around three steps: detecting a vulnerability, verifying that it can be exploited, and then producing a fix.

On SEC-bench Pro, designed to evaluate the discovery of software flaws, Gemini 4 Carbon reaches 88.2%, compared to 83.4% for Argon and 85.4% for GPT-6 Astra.

On ExploitGym, which measures the ability to validate vulnerabilities through exploitation in a test environment, Carbon scores 54.8%, ahead of Argon at 42% and GPT-6 Astra at 42.4%.

Finally, on CWE-bench, dedicated to patching software flaws, the model reaches 71%, compared to 68% for Argon and GPT-6 Astra.

These evaluations were conducted without the safety safeguards typically applied during deployment.

This distinction is crucial. The results describe the model's capabilities in an evaluation environment, not the full range of operations a user will be able to perform with the commercially available version.

They also involve dual-use tasks: the skills useful for verifying and fixing a flaw can also be used in an offensive context.

This increase in capabilities partly explains the cautious approach Google is taking for its deployment. Agents Capable of Continuing Their Work for Several Days

Carbon was not designed solely to answer complex questions.

Google aims to make it a model capable of conducting prolonged technical operations, particularly in large software projects.

The lab claims to have observed Carbon-powered agents pursue tasks for several days, navigate dead ends, search for the root cause of a malfunction, and propose fixes in substantial codebases.

On Terminal-Bench 4.0, Carbon reaches 65.2%, an 8.1-point gain over Argon.

On SWE-sweep, it scores 9.6%, compared to 4.3% for its predecessor.

The gap on SWE-sweep is particularly interesting, as it represents more than double the previous result. The absolute score remains low, however, showing that the tasks covered by this benchmark remain difficult, even for the best models.

In computer use tasks, Carbon also reaches 77.4% on Agents' Last Exam v1.1, compared to 69% for Argon.

Google also reports an 89 Elo point gain on WebDev Arena, where Carbon reaches approximately 1,768 points and fourth place.

These evaluations suggest improvements in autonomy and code quality, without guaranteeing that every long-duration task can be handed over to the model without supervision. Better Utilization of a One-Million-Token Context

The most pronounced results also concern the processing of long sequences of information.

Gemini 4 Carbon supports a context window of up to one million tokens.

On GraphWalks, an evaluation of reasoning over graph structures and contexts ranging between 256,000 and one million tokens, the model reaches 96.2%.

Argon scores 84.2%, GPT-6 Astra 71.8%, and Claude Opus 5.5 66.8%.

The gap is significant: Carbon outperforms Argon by twelve points and Opus by nearly thirty points on this test.

The value of a large context window lies not just in the amount of information a model can receive. It depends on its ability to retrieve relevant relationships and reason correctly over distant elements.

The GraphWalks results suggest progress in this area.

However, they should not be interpreted as a guarantee of reliability across all large documents. Analyzing a code repository, a legal file, or an extended conversation involves different challenges than a graph navigation benchmark. Google Announces Fewer Errors, Shortcuts, and Unauthorized Actions

Agentic capabilities create another challenge: the longer a system works and the more tools it has access to, the