Gemini 4 Carbon strengthens autonomous coding and cybersecurity
Google DeepMind introduced Gemini 4 Carbon, its new frontier model, with a phased rollout initially focused on cybersecurity. The model advances in coding, reasoning, and defense but remains under limited access at launch.
Google DeepMind unveiled Gemini 4 Carbon on October 28, 2026, two weeks after Gemini 4 Argon, with a cautious and targeted launch strategy. This frontier model, described by Google as its strongest to date, improves coding, agentic tasks, and cybersecurity but will not be broadly available immediately. Instead, access begins with a small group of cyber defenders and will proceed through the Fairwind program before reaching developers, enterprises, and then the general public. The message is clear: Google wants to move fast without leaving the door wide open.
Un modèle plus fort, mais pas encore pour tout le monde
Gemini 4 Carbon succeeds Argon and follows the same accelerated release pace. According to Google, the model was designed for long and complex engineering problems, with agentic workflows capable of lasting several days and handling tasks such as root cause analysis or fixing subtle bugs in large codebases. Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, presents it as the company’s strongest model, with top-level cybersecurity and long-term capabilities. However, deployment remains governed by a voluntary pretest process with the US government, indicating that security weighs as much as performance.
In figures published by Google and third-party evaluation platforms, Carbon outperforms Argon on several fronts, but the gap with competitors remains tight. On the new Artificial Analysis Intelligence Index v5, it scores 59 points versus 50 for Argon. On WebDev Arena, it gains 89 Elo points and climbs to fourth place with 1,768 points. The model also progresses on agentic coding tasks, without necessarily dominating everywhere. On Antigravity, it trails Claude Opus 5.5 by three points in agentic coding, reminding us that an AI podium sometimes comes down to just a few lengths.
La cybersécurité devient le terrain de bataille principal
Gemini 4 Carbon is especially highlighted for its cybersecurity performance. Google claims the model better covers the entire defense chain, from vulnerability detection to validation and then correction. On SEC-bench Pro, it achieves 88.2% in autonomous flaw discovery under real conditions. On ExploitGym, it reaches 54.8% in turning a discovered vulnerability into a functional exploit. On CWE-bench, it posts 71.0% for fixing the root cause of flaws. These tests were conducted without safeguards, partly explaining Google’s decision to restrict initial access.
Security does not stop at benchmarks. Google also states that Carbon resisted all attacks tested on Gray Swan, with an attack success rate reduced to 0.0%, compared to 0.7% for Argon. In internal trials, the company claims Carbon produces safe code more than twice as often as Argon and generates quality patches more than ten times better than another competing frontier model, according to the Chrome team. These claims remain partly internal but explain why Google proceeds cautiously.
Accès restreint, prix temporaire et partenaires triés sur le volet
The launch follows a closed-circle logic. Initial access goes to a subset of cybersecurity experts, then to trusted partners via the Fairwind program. Google specifies this program targets priority actors such as governments, critical infrastructure operators, healthcare services, telecoms, and certain technology platforms. Authorized uses remain limited to threat simulation, reverse engineering, or malware analysis for defensive and research purposes. Malicious tasks are prohibited, with background checks and strong authentication constraints enforced.
On pricing, Carbon follows Argon’s launch grid: $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached tokens. After the introductory period, rates rise to $4 and $20 per million tokens. Google has not provided a precise timeline for broader opening but indicates a desire to expand access after feedback. For now, Gemini 4 Carbon remains a cutting-edge model mainly visible from the specialists’ showcase.
By the numbers
- 59 points — Gemini 4 Carbon’s score on the Artificial Analysis Intelligence Index v5.
- 88.2% — result on SEC-bench Pro, according to Google.
- 54.8% — score on ExploitGym, according to Google.
- 71.0% — performance on CWE-bench, according to Google.
- $2 / $10 — launch price per million tokens, input/output.