Gemini 4 Carbon release notes

Google has officially released Gemini 4 Carbon, its most capable model to date, positioned for high-stakes knowledge work, agentic software engineering, and scientific discovery. Carbon is designed to handle tasks that require deep reasoning and extended verification loops, distinguishing it from the previously released Gemini 4 Argon, which remains the primary model for general software and knowledge tasks.

Capabilities and Performance

Gemini 4 Carbon is optimized for complex, multi-step workflows. According to Google, the model excels in:

  • Agentic Coding: It supports multi-package migrations, cross-subsystem root-cause analysis, and full-stack development. It is being integrated into Google Antigravity with a new /swarm feature, which coordinates multiple agents to handle architectural design and adversarial verification.
  • Cybersecurity: Carbon provides advanced vulnerability detection and patching. In testing, it demonstrated high precision in triage and recall, and has been utilized by researchers to identify 0-day vulnerabilities in open-source software.
  • Scientific Reasoning: The model shows improved performance in physics, mathematics, and life sciences. It has been used to assist in formal verification of cryptographic protocols and to accelerate protein structure prediction by discovering remote homologs in large sequence databases.

Carbon is priced at $2 per million input tokens and $10 per million output tokens, matching the introductory pricing of Gemini 4 Argon.

Access and Availability

Access to Gemini 4 Carbon is currently tiered based on user type:

  • Developers and Enterprises: Available via the Gemini API in Google AI Studio and Google Cloud’s Gemini for business.
  • Consumers: Rolling out to Google AI Ultra subscribers within the Gemini app and AI Mode in Search.
  • Scientists and Researchers: Eligible participants can apply for access to Carbon through the Google Antigravity research program, which provides six months of free access to frontier models.
  • Cybersecurity Defenders: Access is managed through the Fairwind Program.

Community and Industry Reception

Initial community feedback on Hacker News highlights a mix of interest in the model's performance and frustration regarding Google's tiered access model. Some users reported that the model provides high-quality, nuanced responses that outperform previous iterations, particularly in complex research tasks.

However, users expressed concern over the complexity of Google's subscription tiers, with some noting that the "Ultra" plan requirement creates a barrier to entry compared to competitors. Others pointed out that while the model is competitive, the delay in its release relative to other industry-leading models like Fable 5.2 has impacted its perceived market position. Technical users also noted that the model's performance in agentic tasks, particularly when paired with multi-agent frameworks like /swarm, represents a significant step forward for automated engineering workflows, though some users reported latency issues in the current implementation.

Sources