Gemini 4 Carbon Takes the Lead in AI Benchmarks

Google DeepMind presented Gemini 4 Carbon, two weeks after Argon. The model climbs to the top of several benchmarks, prioritizing autonomous development uses and cybersecurity.

Google DeepMind launched Gemini 4 Carbon on October 28, 2026, two weeks after Gemini 4 Argon, with a clear goal: to push further autonomous workflows over several days, notably in software development and security. On the Artificial Analysis Intelligence Index v5, the model climbs to 59 points compared to 50 for Argon, while remaining neck and neck with the best systems on the market. Access is initially limited, first to cybersecurity experts, as Google submits Carbon to a voluntary pre-test by the US government. Not exactly the kind of release entrusted to just anyone.

A model designed to last several days on the same project

Google presents Gemini 4 Carbon as a model designed for "much harder" and long-term engineering problems. In its internal tests, the AI worked several consecutive days on complex tasks, ranging from root cause analysis to fixing subtle bugs in large code repositories and low-level software.

The company also says it has improved scientific reasoning, multimodal understanding, and full-stack web development based on user feedback. On WebDev Arena, Carbon gains 89 Elo points and reaches 1,769 points, placing it in fourth position. On the agentic coding side, it narrows the gap with Claude Opus 5.5 but remains behind the latter on some specialized rankings.

Security becomes a product feature, not just a safeguard

The cybersecurity aspect occupies a central place in this version. Google claims that Gemini 4 Carbon covers the entire defense cycle, from vulnerability detection to patch writing. On Gray Swan, the success rate of indirect prompt injection attacks drops from 0.7% with Argon to 0.0% with Carbon, according to figures provided by Google. This does not mean prompt injection is completely solved, but the test seems to have reached its limit.

On security benchmarks, the results are even more decisive. SEC-bench Pro shows a score of 88.2%, CWE-Bench 71.0%, and ExploitGym 54.8% after ten hours of execution. Google also specifies that some tests were conducted without security safeguards, which partly explains the very controlled framing of the launch. Cybersecurity experts are served first; others will wait their turn.

Close rankings, but contained cost

Independent benchmarks confirm a clear improvement compared to Argon. On Artificial Analysis, Gemini 4 Carbon takes the lead of the Intelligence Index v5 with 59 points, against 50 for its predecessor. On agentic knowledge tasks, it gains 126 Elo points to reach 1,609, without surpassing some competitors like Kimi K3.1 Pro or Muse Wonder.

However, not everything progresses at the same pace. The correct response rate rises from 48% to 58%, but the hallucination rate climbs from 4% to 11%, while remaining among the lowest in the ranking. On Vals AI, Carbon reaches 70.36% accuracy, behind Claude Sonnet 5.5 at 70.85%, but ahead of Claude Opus 5.5 at 70.18% and Argon at 69.24%.

The strong point here lies in the performance-to-cost ratio. Google maintains the same launch price as Argon, namely 2 dollars per million input tokens and 10 dollars per million output tokens. According to Artificial Analysis, the cost per task remains close to that of Argon, around 7.76 dollars, for a higher level of performance.

What this launch says about Google's strategy

This two-week publication pace shows that Google wants to occupy the field very quickly, without giving competitors time to lock in comparisons. The strategy is clear: push a stronger model on long tasks, better positioned on security, and competitively priced enough to remain attractive to developers and businesses. Between speed of release and regulatory caution, Gemini 4 Carbon mainly embodies a bet: to go faster than the market without burning its wings along the way.

Google then plans to open the model to developers, businesses, and later to the general public via its product interfaces, once feedback is collected and safeguards adjusted. At this stage, Carbon mainly establishes itself as a transition model towards more autonomous uses, where AI no longer just responds quickly but must also endure.

To read