- Published on
Just In: Gemini 4 Carbon Officially Released, Hachimi Returns to World No. 1
- Authors
- Name
- DeepHorus
Google DeepMind has officially released Gemini 4 Carbon. The company says it scored 59 on the Artificial Analysis Intelligence Index v5.0, surpassing Claude Opus 5.5's 58 and GPT-6 Astra's 57 to top the global rankings. According to Arena.ai, it ranks first in Text Arena, second in Agent Arena, and fourth in Code Arena WebDev.
This release is seen as Google's response to competitive pressure. Previously, Google's flagship Pro model had long failed to materialize, while Anthropic and OpenAI continued to raise the ceiling of capabilities. Gemini 4 Argon only debuted two weeks ago, and Carbon's composite index is 9 points higher than Argon's. Yao Shunyu, who moved from Anthropic to Google DeepMind, also promoted Gemini.
Carbon's standout capabilities are in long-horizon engineering and cybersecurity. Google says it can run continuously for several days to handle large codebases and low-level system problems, and when blocked, continue searching for solutions and fix hidden defects. On WebDev Arena, its Elo increased by 89 points to 1,769, ranking fourth, and it is positioned as an autonomous agent.
In cybersecurity tests, Carbon achieved 88.2%, 54.8%, and 71.0% on SEC-bench Pro, ExploitGym, and CWE-bench, respectively. Google says the results were measured without safety guardrail restrictions enabled, covering vulnerability discovery, risk validation, and generation of remediation plans. The Chrome team says its security patch performance is more than ten times higher than that of another lab's frontier model.
In Terminal-Bench 4.0, Carbon's successful cheating rate was 0.0% and attempted cheating rate was 0.4%; high-severity behavior alerts decreased by nearly 80%, while unauthorized actions and insufficient completion decreased by 26% and 29%, respectively. In Gray Swan's indirect prompt injection test, the attack success rate dropped from 0.7% to 0.0%.
General users cannot use it for the time being. Google says it will first provide it to some cybersecurity defenders this week, then expand it to trusted partners through the Fairwind Program, and participate in the U.S. government's pre-release model testing; there is no timetable for opening it to general users. Pricing follows Argon's launch pricing: $2 per million input tokens and $10 per million output tokens, adjusted to $4 and $20 after the promotional period.
The release sparked discussion on Hacker News. Some users disliked that Google announced it early but delayed opening access, while others said they could not be bothered to click the release link. The 'performance without the actual object' style of release has drawn attention again; for Gemini, letting more users actually use the global No. 1 may be more important than announcing another top ranking.
