📰 Key Takeaways

Gemini 4 Carbon is Google DeepMind’s new flagship model, built around state-of-the-art cybersecurity capabilities and long-horizon task handling. It ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59, with big improvements in coding and agentic tasks. On security, Carbon scores 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploiting them), and 71.0% on CWE-bench (patching them), all well ahead of the pack. In internal testing, its success rate at producing secure code is more than double that of Gemini 4 Argon, which launched two weeks ago, and it also shows less misalignment and fabrication in real coding environments.

For long-horizon engineering work, Carbon is designed for agent workflows that keep running. The team saw it work continuously for days, keep trying after hitting dead ends, and tackle subtle bugs in large codebases and low-level systems, including finding the root cause and patching it. Web development quality is up too: it jumped 89 Elo points on WebDev Arena to 1768, ranking fourth.

The rollout is deliberately cautious and staged. Google is taking part in the US government’s voluntary pre-release model testing. Starting this week, access goes first to a group of cyber defenders, then expands to more trusted partners through the Fairwind Program. Google will keep collecting feedback and tuning safeguards, then open it up to developers, enterprises, and general users as quickly as it can. Pricing is the same as Argon: $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens. The later part of the original post, covering hands-on reports from the Chrome team, was truncated, so see the original link for details.

💬 JudyAI Lab Take

Gemini 4 Carbon puts cybersecurity front and center for a flagship model. It leads by a wide margin on all three benchmarks (finding, exploiting, and patching vulnerabilities), and it also takes first place on the Artificial Analysis Intelligence Index v5.0 with a score of 59.

What this news shows is that the competition between models is shifting from “how smart are the answers” to “can it do real work reliably over a long time.” The original post says Carbon is built for agent workflows that keep running: it can work for days, keep going after hitting a dead end, and find the root cause of subtle bugs in large codebases and low-level systems before patching them. The other thing worth noticing is the release approach. Google is joining the US government’s voluntary pre-release testing, opening access to a group of cyber defenders first, expanding through the Fairwind Program, and only then opening to general users. The stronger the capability, the more it makes sense to put it in defenders’ hands first and widen access later, and this kind of staged release may become the norm for highly capable models. Pricing matches Argon at $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input, which means the cost barrier for long-running agent tasks didn’t go up even though the capability did.

My suggestion: list the most time-consuming code review or debugging workflows you have, think about which part could be handed to an agent that can run for a long time, and leave room for a human review step.

📅 Source Info

🔗 Further Reading