Gemini 4 Carbon pushes Google towards cybersecurity and coding uses

Google unveiled Gemini 4 Carbon, an AI model focused on cybersecurity and long programming tasks. Access will initially be limited to selected defenders before wider deployment.

Google introduced Gemini 4 Carbon on October 28, describing it as its most advanced AI model to date for cybersecurity and long coding tasks. The system is designed to detect vulnerabilities, assess risks, and propose fixes, with access initially reserved for a select group of cybersecurity defenders. For Google, the challenge goes beyond simple benchmark rankings: it is about demonstrating that a model can work on engineering problems that sometimes take several days.

Un modèle pensé pour casser des bugs, pas seulement des records

Gemini 4 Carbon continues the legacy of Gemini 4 Argon, launched two weeks earlier according to Google. Koray Kavukcuoglu, Senior Vice President at Google DeepMind and Chief AI Architect at Google, stated that Carbon also improves coding, scientific tasks, web development, and AI agents capable of chaining multiple work steps. In short, the model is judged not only on its speed but on its ability to sustain performance when projects pile up.

Google specifies that Carbon targets complex engineering problems, such as finding the origin of bugs and then fixing them in large codebases and low-level systems. The company also claims that its external cybersecurity evaluations were conducted without safeguards to measure the model's behavior under more realistic conditions. The kind of test that does not forgive approximations, even for the most talkative models.

Des scores élevés, mais encore peu de monde pour les vérifier

On paper, the numbers are solid. Google says Gemini 4 Carbon scored 88.2% on SEC-bench Pro, a benchmark dedicated to identifying vulnerabilities in real software. The model also achieved 54.8% on ExploitGym, which evaluates exploit development, and 71.0% on CWE-bench, focused on fixing software weaknesses. The company adds that its internal tests show a success rate more than twice that of Argon for generating secure code.

Chrome, at Google, also reportedly observed better performance than another leading AI model, with a gap of more than ten times in generating quality patches. However, Google did not name the competitor nor detail the comparison method [to be verified]. On Artificial Analysis, Carbon tops the Intelligence Index v5.0 with 59 points, while Google attributes an 89-point gain on WebDev Arena, where it ranks fourth with 1,769 points. Again, caution is advised until public access is opened.

Accès limité, tarification basse au lancement

Deployment will begin with selected cybersecurity defenders, then expand to other trusted partners via the Fairwind Program. Google also indicates participation in the US government's voluntary process to test models before release. No precise date has been given for wider public access, but the company says it wants to proceed in phases to consolidate safeguards.

On pricing, Gemini 4 Carbon is offered at an introductory rate of $2 per million input tokens and $10 per million output tokens, the same prices as Gemini 4 Argon at launch. After this period, rates will rise to $4 and $20 per million tokens. Cached tokens receive a 95% discount on the standard input price. For developers, economic calculation matters almost as much as benchmarks.

Ce que Google essaie de reprendre aux autres

Google spent much of 2026 chasing its rivals in generative AI, according to The Rundown, before returning to the lead with two models launched within weeks. Carbon is therefore meant as both a technical demonstration and a product: to show that Google can still compete on cutting-edge models while targeting uses where reliability is as important as speed. It remains to be seen if this position will hold when the model leaves the restricted circle of testers. The AI podium changes quickly, and no one keeps their medal warm for long.

By the numbers

  • October 28, 2026 — date of Gemini 4 Carbon presentation, according to Google.
  • 88.2% — score on SEC-bench Pro, vulnerability benchmark.
  • $2 / $10 — introductory price per million input/output tokens.
  • 95% — discount on cached input tokens.
  • 1,769 points — score announced on WebDev Arena by Google.

To read