Back to blog
Google Built Gemini 4 Carbon to Hunt Bugs Before Attackers Do. Here's the Fine Print

Google Built Gemini 4 Carbon to Hunt Bugs Before Attackers Do. Here's the Fine Print

October 30, 2026
By Vamsi Reddy and Sai Goutham Reddy

Google Built Gemini 4 Carbon to Hunt Bugs Before Attackers Do. Here's the Fine Print

Google DeepMind announced Gemini 4 Carbon on October 28, and it isn't pitched as a better chatbot. Google calls it our strongest model with state-of-the-art cybersecurity and long horizon capabilities. Put simply, it's built to find security holes, check that they're real, and patch them.

The other half of the pitch is stamina. Google says teams have watched Carbon run for days on multi-step engineering tasks. That's a very different job from autocomplete.

There's a catch, and it's a big one. Almost nobody can use it yet.

Key Features and the Numbers Google Is Claiming

All of these are Google's own figures. Google also notes that the cybersecurity scores were evaluated without guardrails, so the version you eventually get may score lower.

  • Vulnerability discovery: 88.2% on SEC-bench Pro.
  • Exploitation: 54.8% on ExploitGym. Yes, a defense model gets graded on breaking things too. You can't validate a bug you can't trigger.
  • Patching: 71.0% on CWE-bench, per Collinear's leaderboard.
  • General intelligence: a score of 59 on the Artificial Analysis Intelligence Index v5.0, which Google says ranks first.
  • Agentic coding: 65.2% on Terminal-Bench 4.0 and 61.2% on FrontierSWE 2. That second one's worth flagging. OpenAI's GPT-6 Astra leads it at 65.5%, according to DeepMind's own comparison table.
  • Web dev: 1768 on WebDev Arena. That's fourth place, and Google says it's an 89-point Elo gain.
  • Safer output: Google reports more than 2x better secure-code generation than Gemini 4 Argon on internal tests. It also claims a 0.0% attack success rate on Gray Swan's indirect prompt-injection benchmark.

Pricing. It launches at the same introductory rate as Argon: $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After the intro period that doubles to $4 and $20. Google hasn't said when the intro period ends.

Real-World Use Cases

Continuous vulnerability triage. This is the headline job. Point Carbon at a codebase, let it find candidate bugs, confirm they're exploitable, then draft the fix. Security teams buried in scanner false positives have the most to gain here.

Long-running refactors. If the run for days claim holds up, Carbon fits migrations that wear down shorter-lived agents. Think dependency upgrades across dozens of services, or slowly moving a legacy module onto a new API.

Agents that read untrusted input. A near-zero prompt-injection rate matters most when your agent reads emails, tickets, or web pages written by strangers. That's where injection attacks actually happen. One benchmark isn't a guarantee, but it's the right thing to optimize for.

Cost-sensitive agent loops. Cheap cached input means tokens you resend in every turn, like repo context and system prompts, cost a fraction of the normal rate. Long agent loops resend a lot, so the discount adds up fast.

Hard Limitations

You probably can't get it. Google says Carbon is going to a subset of cyber defenders first. Next come additional trusted partners through its Fairwind Program. Developers, businesses and consumers are coming soon, with no date.

Self-reported benchmarks. Every number above comes from Google's launch post. The headline cyber scores were run without guardrails, and nobody outside Google has replicated them yet.

It doesn't win everything. GPT-6 Astra beats it on FrontierSWE 2, and fourth place on WebDev Arena isn't first. If your work is mostly front-end, Carbon isn't automatically the pick.

The price doubles later. Budget for $4/$20, not $2/$10. A multi-day agent run burns a lot of output tokens, and output is the expensive side.

Dual-use is real. A model that scores 54.8% on an exploitation benchmark is useful to attackers too. That's why access is gated. Google says it's taking part in the U.S. government's voluntary pre-release testing process. Expect identity checks and usage policies when access widens.

Should You Wait for It?

If you run a security team, apply for early access now and line up a test repo full of known CVEs. That's the quickest way to check Google's numbers against your own code.

Everyone else can carry on with what they're using today. Carbon's pricing and the days-long agent pitch are worth watching, but a model you can't call doesn't ship anything. Check back once Google gives a real date for developer access.