Gemini 4 Carbon release notes

Overview of Gemini 4 Carbon

Gemini 4 Carbon is Google DeepMind's latest frontier model designed for high-precision knowledge work, complex software engineering, and scientific discovery. While the previously released Gemini 4 Argon remains the standard for general software engineering and knowledge tasks, Carbon is engineered for deep reasoning, multi-step agentic workflows, and extended verification loops. It is now available to developers, enterprises, and consumers via the Google AI API, Google Antigravity, the Gemini app, and AI Mode in Search.

Agentic Capabilities and Performance

Carbon is optimized for complex, multi-agent environments. Google is previewing /swarm in Google Antigravity, a feature that allows users to turn a single prompt into a coordinated team of agents. This enables developers to assign Carbon to high-level roles like architectural design and auditing, while offloading parallel tasks to lighter models.

Carbon demonstrates strong performance across several industry benchmarks:

  • Coding and Automation: Ranks #1 on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench.
  • Multimodal Knowledge: Ranks #2 on the Multimodal Knowledge Agent (MKA) benchmark, with a 6X lower cost per task compared to the top-performing model.
  • Financial Research: Matches top-tier performance on Hebbia’s E2E Financial Agentic evaluation and ranks at the top of Balyasny Asset Management’s General Financial Intelligence benchmark.

Cybersecurity and Vulnerability Research

Carbon introduces advanced cybersecurity capabilities, including vulnerability detection, exploitation, and patching. The model has demonstrated significant improvements in security-focused benchmarks:

  • SEC-bench Pro, ExploitGym, and CWE-bench v1: Leads in finding, exploiting, and patching vulnerabilities.
  • Project Fortify: Achieved 100% recall on scanning and 95% precision on triage in internal benchmarks.
  • Real-world Impact: Used by Google’s Cloud Vulnerability Research team and Wiz to identify and remediate critical 0-day vulnerabilities in open-source software, in some cases compressing weeks of research into a single afternoon.

Scientific and Mathematical Reasoning

Carbon enhances scientific discovery by synthesizing experimental data and testing hypotheses. Key applications include:

  • Formal Verification: Used within the AlphaProof Code framework to mathematically prove security properties of the Signal cryptographic protocol, reducing a task that historically took years to weeks.
  • Assay Development: Diagnosed subtle experimental anomalies in wet lab data, such as incomplete automated mixing, to improve reproducibility.
  • Protein Structure Prediction: Increased the retrieval of remote homologs for understudied protein families by 40-fold on average, providing more data for structure prediction models like AlphaFold 3.

Safety and Security Mitigations

Reflecting its increased capabilities, Carbon includes a strengthened safety stack designed to prevent misuse and prompt injection.

  • Attack Mitigation: In simulated cyberattack environments, Carbon blocked 99% of attack trajectories from adaptive red-teaming agents. The per-attempt breach rate was reduced to 0.8%, down from 3.1% in Argon.
  • Honesty and Alignment: On a 375-task agentic honesty benchmark, Carbon passed 92% of tasks, compared to 70% for Argon. It demonstrated a 98% rate of declining to fill in missing experimental data.
  • Prompt Injection: Carbon achieved a 0% success rate on the Gray Swan indirect prompt injection (IPI) benchmark.

Sources