Google DeepMind announces Gemini 4 Carbon, starting with cyber defenders

Google says Carbon ranks first on the recently updated version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59[1]. The Decoder reports that Argon sits at 50 on the same index, and that on Artificial Analysis's Coding Agent Index Carbon's 59 still trails Claude Opus 5.5 at 62[3]. The Verge notes that Claude Opus 5.5 and GPT-6 Astra trail Carbon on the Intelligence Index[2].

On web development, Google says Carbon gained 89 Elo points to reach fourth place on WebDev Arena at 1768[1]. It also says the model is built for long agentic work, and that its own teams have seen Carbon "run for days" through dead ends on tasks such as root-causing and patching subtle bugs in large codebases and low-level systems[1].

The biggest claims are about security. Google reports, measured without guardrails, 88.2% on SEC-bench Pro for finding real vulnerabilities, a state-of-the-art 54.8% on the exploitation benchmark ExploitGym, and 71.0% on Collinear's CWE-bench for turning findings into working fixes[1]. On internal benchmarks it says Carbon more than doubles Argon's success rate at generating secure code, and that the Chrome team found it more than 10 times better than "another lab's large frontier model" at producing high-quality patches[1]. Wiz, Google's cloud security platform, also sees Carbon leading on web penetration testing benchmarks, according to the post[1].

Google says Carbon uses the same layered safety stack as Argon, including real-time misuse defences, prompt-injection resilience and chain-of-thought monitoring, and was trained to handle high-pressure tasks better[1]. Its figures include successful cheating on Terminal-Bench 4.0 cut to 0.0% with cheating attempts at 0.4%, high-severity misalignment alerts in internal live coding down by almost 80%, unsupported claims roughly halved compared with Argon, and a 0.0% attack success rate on Gray Swan's Indirect Prompt Injection benchmark, down from Argon's 0.7%[1]. It also says overreach fell by 26% and under-service by 29% on multi-step plans[1].

Google says it is "actively engaged in the U.S. government's voluntary process for pre-release model testing"[1]. Access starts this week with a subset of cyber defenders, then widens to more trusted partners through its Fairwind Program, and reaches developers, businesses and consumers "as soon as possible", with no date given[1].

Pricing will start at Argon's introductory rate of $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper, then rise to $4 and $20 after the introductory period ends[1].

Carbon follows the same pattern as other recent frontier launches: the strongest security capabilities go to defenders before anyone else, while the lab tunes guardrails. For teams building coding agents, the practical points are the multi-day task claims and the unchanged introductory price. Until Carbon reaches the Gemini API, though, most of its numbers are Google's own, and independent testing will have to wait for wider access[1].