On October 29, Google DeepMind officially announced the launch of its new-generation flagship large model — Gemini 4 Carbon. In the latest update of the Intelligence Index v5.0 evaluation rankings by the authoritative independent organization Artificial Analysis, the model achieved a high score of 59, surpassing the previously long-dominant Claude Opus 5.5 by 1 point and leading GPT-6 Astra by 2 points. Compared with the previous generation released two weeks ago, Gemini 4 Carbon jumped directly by 9 points, successfully reclaiming the global No. 1 throne in overall large model capability. Meanwhile, in Arena.ai's latest blind test rankings, Gemini 4 Carbon also delivered impressive results, taking first place in Text Arena and second place in Agent Arena.
In terms of core capabilities, Gemini 4 Carbon has locked its evolution focus on two hardcore directions: "Long-horizon Engineering" and "cybersecurity". In handling long-horizon tasks, the model demonstrates extremely strong sustained operational capability, able to run efficiently for several consecutive days, deeply process very large codebases and complex low-level system architectures, autonomously locate and repair hidden software defects that are normally extremely hard to detect, and firmly rank in the industry's top tier across multiple coding and agent benchmarks.
In the cybersecurity field, the model has achieved end-to-end integration from vulnerability discovery and risk verification to code repair. It performed exceptionally well in multiple professional security evaluations: SEC-bench Pro scored 88.2%, ExploitGym scored 54.8%, and CWE-bench scored 71.0%. Official data shows that its success rate in secure code generation is more than twice that of the previous-generation product Argon. At the same time, the model's safety and reliability have also been upgraded in step: in Terminal-Bench 4.0 testing, its cheating success rate dropped to 0%, high-severity behavioral inconsistency alerts decreased by nearly 80%, and its defense success rate against indirect prompt injection attacks reached 100%, with significant improvements also seen in unauthorized actions and incomplete task completion.
In terms of release pace and commercial pricing, Gemini 4 Carbon adopts a steadily advancing strategy. This week, the model will first be opened to a small number of cybersecurity defense professionals, followed by a gradual expansion of testing capacity through exclusive partnership programs, while the timeline for full availability to ordinary users and developers remains unclear. On pricing, the model follows Argon's launch standard: during the promotional period, it costs $2 per million input tokens and $10 per million output tokens, with the official price doubling after the promotional period ends.
This high-frequency release cadence has also sparked varied discussions in the industry. Many observers believe that the current model of top large model vendors frequently releasing new models while delaying full public availability leans to some extent toward intense leaderboard ranking competition, and their true practical deployment value and everyday experience still await deep testing by developers and users.