Google DeepMind announced Gemini 4 Carbon on Monday, calling it the company's strongest model to date and built around cybersecurity work and tasks that require reasoning over long stretches of time. According to DeepMind, access starts this week, but only for a restricted group of cyber defense teams.

The company says Carbon ranks first on the newly updated v5.0 Artificial Analysis Intelligence Index, with a score of 59. On internal cybersecurity benchmarks cited by DeepMind, the model scored 88.2% on SEC-bench Pro (finding vulnerabilities in code), 54.8% on ExploitGym (exploiting those flaws) and 71% on CWE-bench (patching them). Because those figures come only from the company itself, they describe what DeepMind reports about the model, not an independent measurement.

Carbon arrives two weeks after Gemini 4 Argon, which launched in October, and keeps Argon's introductory pricing: $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens, according to DeepMind.

Built for long engineering jobs, not quick answers

Unlike a model tuned for fast replies, Carbon was designed for engineering problems that take a long time to work through. DeepMind says that in internal tests, the model ran for days at a stretch to push past dead ends on tasks such as tracking down and patching specific bugs across large codebases and low-level systems.

According to the company, Carbon also jumped 89 Elo points on WebDev Arena, a leaderboard that scores the quality of web interfaces generated by AI models, landing it in fourth place overall with 1,768 points. DeepMind says the model also achieved more than double the success rate at generating secure code compared with Argon, on its own internal benchmarks.

DeepMind also says Google's own Chrome team reported Carbon performing more than ten times better than another, unnamed lab's large frontier model at generating high-quality patches. Wiz, the cloud security platform owned by Google, says it separately observed Carbon leading on benchmarks for web penetration testing, used to check for security flaws before real attackers can exploit them. Because Wiz belongs to the same corporate group as Google, that observation doesn't count as a second independent source — it's another claim relayed by DeepMind about its own model.

A deliberately slow rollout

Carbon isn't going straight to the public. DeepMind describes the rollout as intentionally gradual: initial access goes only to a group of cyber defenders, then to additional trusted partners through what the company calls the Fairwind Program. There is no date yet for opening the model to developers, businesses or consumers through Google's own products.

The company also says it is taking part in the U.S. government's voluntary process for pre-release model testing. According to DeepMind, Carbon runs on the same safety stack as Argon — combining training-time safeguards, real-time misuse defenses and monitoring of the model's chain of thought — with additional training aimed at keeping the model's responses appropriate under high-pressure tasks.

Chain-of-thought monitoring, according to DeepMind, works as an audit of the path the model takes to reach an answer, not just of the final output shown to the user. The company says that check helps catch the model drifting from expected behavior before it shows up in the response it gives — a safeguard DeepMind describes as more important for a model built to run for days at a time without constant human supervision.

In practice

  • Who can use it: for now, only a restricted group of cyber defenders and Fairwind Program partners, per DeepMind; developers, businesses and consumers do not yet have access through Google's products.
  • Since when: access for that first group began this week, according to the company.
  • What it costs: $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached tokens — the same introductory price as Gemini 4 Argon, per DeepMind.
  • What changes from Argon: the company reports gains in secure code generation, long-horizon engineering tasks and web development quality, while keeping Argon's launch pricing.