Google Drops Gemini 4 Carbon: A Frontier AI Built for Cyber Defense
Google just launched Gemini 4 Carbon. It runs for days to fix bugs and dominates security benchmarks, but you cannot use it yet.
Models · Source: Hacker News
What happened
Google DeepMind just announced Gemini 4 Carbon. It comes only two weeks after they dropped Gemini 4 Argon. This is their newest frontier model. It focuses heavily on cybersecurity and long horizon tasks. Carbon ranks first on the new Artificial Analysis Intelligence Index with a score of 59.
Carbon is built for sustained agentic workflows. Google says internal teams let the model run for days to tackle complex engineering problems. It digs through large codebases, finds root causes, and patches subtle bugs. It also jumps 89 elo points to land at number four on the WebDev Arena leaderboard. The Chrome team reportedly saw it perform ten times better than a rival model at generating high quality patches.
You cannot access it yet. Google is rolling it out slowly. They are giving early access to a select group of cyber defenders through their Fairwind Program. They are also doing voluntary pre-release testing with the US government. When it does launch, it will cost two dollars per million input tokens and ten dollars per million output tokens.
Key facts
- 59 — Score on the Artificial Analysis Intelligence Index v5.0, ranking it number one.
- 88.2% — Score on SEC-bench Pro for autonomously finding real-world vulnerabilities.
- 0.0% — Indirect Prompt Injection attack success rate on the Gray Swan benchmark.
- $2 — Introductory cost per million input tokens before caching discounts.
- 95% — Discount applied to cached input tokens.
Why it matters
The era of quick prompt and response is shifting to long running agents. If Carbon can actually run for days without going off the rails, it changes how we build software. You will not just ask an AI to write a single function. You will assign it a complex ticket and check back later. The model also boasts a zero percent success rate for indirect prompt injections. This is a massive deal for building secure enterprise apps. It means agents can handle untrusted data without getting hijacked.
Security is becoming the main battlefield for frontier models. Google is explicitly targeting vulnerability detection, validation, and remediation. This puts immense pressure on dedicated security startups. If a general frontier model can autonomously uncover flaws and write working patches better than specialized tools, the market for thin wrapper security products will collapse. Google is training these models to refuse shortcuts and stop fabricating evidence. They want enterprise trust.
For builders
Cheap cached tokens for massive codebases
Carbon keeps the introductory pricing of Argon. You pay two dollars per million input tokens, but cached inputs get a massive 95 percent discount. Builders who load massive repositories into context will save a fortune.
Zero percent prompt injection risk
Google claims Carbon completely resisted indirect prompt injections on the Gray Swan benchmark. If this holds up in production, enterprise developers can finally trust AI agents with untrusted user data. You spend less time building guardrails and more time shipping features.
Threat to specialized security startups
Carbon scores highly on finding vulnerabilities and writing patches. Founders building AI security tools now compete directly with the base model from Google. If you just wrap an LLM to find bugs, you will lose your customers to this native capability.
My take
Google is moving insanely fast. Shipping Carbon two weeks after Argon shows they are done playing games. I love that they are focusing on long running agents that actually finish tasks instead of just chatting. But gating it behind government testing is frustrating for builders who want to test these limits today. We need access to build.
Original reporting: Hacker News. This is my rewrite and opinion.