SIAgentNews.com
← All news

Google launches Gemini 4 Carbon for complex reasoning and cyber tasks

The new model offers frontier performance in scientific research and cybersecurity, featuring a coordinated multi-agent system for complex workflows.

· First reported by deepmind.google

0.8%Carbon's per-attempt breach rate in cyberattack testing

The short version

  • Carbon ranks #1 on CursorBench 4.0 and Zapier’s AutomationBench at over 2X lower cost per task than the next best model.
  • The model achieved 100% recall on scanning and 95% precision on triage in internal security testing via Project Fortify.
  • Researchers used Carbon to help prove the cryptographic protocol of Signal and discover novel drug toxicity mechanisms.
  • It is available to paying developers and Google AI Ultra subscribers starting today at the same price as Gemini 4 Argon.

Why it matters

For users, this means a shift from simple chat to 'agentic' workflows where the AI can manage entire projects. It specifically targets high-stakes industries like drug discovery and cybersecurity where reasoning accuracy is critical.

A new tier of intelligence

Google is releasing Gemini 4 Carbon, a model designed for tasks that push the limits of artificial intelligence. While the previous Gemini 4 Argon model is suited for standard software engineering and knowledge work, Carbon is built for deeper reasoning. It is specifically intended for open-ended scientific research, complex multi-step coding, and debugging.

The model uses more tokens than Argon to explore information thoroughly and run extended verification loops. This means it can reason more deeply over large datasets or complex instructions. It aims to provide frontier performance for users who need the AI to handle high-complexity knowledge work that requires more than a single-step response.

Carbon is now available to developers, enterprises, and consumers. It follows the launch of Argon and is positioned as the strongest model in the Gemini 4 family. It is available via the API in Google AI Studio and Gemini for business for all paying users at an introductory price of $2 per million input tokens and $10 per million output tokens.

Coordinated agentic workflows

One of the most significant features of Carbon is the preview of /swarm within Google Antigravity. This tool allows a single prompt to turn into a coordinated team of AI agents that users can track in real time. Instead of one model doing everything, /swarm lets users assign Carbon to specific roles that demand deep reasoning, such as architectural design or auditing.

This system can pair Carbon with faster, lighter models to handle parallel tasks in isolated workspaces. It also uses adversarial verifiers to test every change made by the agents. This multi-agent approach is particularly effective for engineering tasks like multi-package migrations and cross-subsystem root-cause investigations.

For web app development, the model can run build-and-verify loops to produce complete applications. It checks both the underlying code and the rendered user interface to ensure the final product works. This capability allows for more autonomous development cycles where the AI can recover from errors and verify its own output.

Scientific and cyber breakthroughs

The model has already shown significant results in specialized fields like drug discovery and cybersecurity. For the company Axiom Bio, Carbon became the leading model for translational science by finding novel mechanisms of toxicity. It also helped researchers at Google accelerate formal verification by generating millions of lines of code in Lean to prove the security of the Signal cryptographic protocol.

In cybersecurity, Carbon provides end-to-end intelligence for protecting software. When tested in Project Fortify, it hit 100% recall on scanning and 95% precision on triage. It also helped the Cloud Vulnerability Research team prove that two bugs in a popular open-source software were exploitable, leading to a swift patch from the maintainers.

The model also assists in protein structure prediction by discovering remote homologs in sequence archives. In a blind test, Carbon increased the number of retrieved homologs by 40-fold on average for certain protein families. This helps provide more information for structure prediction models without requiring them to be retrained.

Safety and honesty benchmarks

Google is emphasizing that Carbon includes stronger safety and security mitigations to prevent misuse. In simulated cyberattack testing, the model's safeguards blocked 99% of attack trajectories from an adaptive red-teaming agent. This lowered the per-attempt breach rate to 0.8%, which was a significant improvement over the 3.1% rate seen on the Argon model.

The model also shows improved alignment and a reduction in data fabrication. On a 375-task agentic honesty benchmark, Carbon passed 92% of tasks, compared to 70% for Argon. It declined to fill in missing experimental data 98% of the time and flagged methodological flaws in 68% of the cases.

Additionally, Carbon achieved a 0% success rate on the Gray Swan indirect prompt injection benchmark. This marks the first time a model has reached that milestone. However, Google notes that prompt injection robustness remains an area for continued improvement as they continue to test against sophisticated attacks.

Availability and access

Gemini 4 Carbon is rolling out across several Google platforms today. Google AI Ultra subscribers can access it in the Gemini app, AI Mode in Search, and Google Antigravity. Paying developers and enterprises can access the model through the API in Google AI Studio, Android Studio, and Gemini for business.

For the scientific community, Google is expanding access to Carbon for participants in a special program. This program provides free access to both Argon and Carbon for six months to help accelerate research. Applications for this program are currently being accepted on a weekly rolling basis based on eligibility and capacity.

Cybersecurity defenders can also apply to join the Fairwind Program for access to Carbon for cyber. This program provides the intelligence needed to protect critical software. Developers can find more information in the model card and updated developer guides to begin migrating their existing workflows to the new model.

What to watch

  • How the /swarm multi-agent coordination performs on large-scale enterprise architectural tasks.
  • The results of the responsible disclosure for the open-source software vulnerabilities found by the CVR team.

Try it

If you are a Google AI Ultra subscriber, check the Gemini app or AI Mode in Search to see if Carbon is active for your queries.

Sources and links

First reported by deepmind.google. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news