The model succeeds Gemini 4 Argon, with improvements in vulnerability detection, automated patching, and reliability. Initial access begins this week through Google's Fairwind Program, with broader availability planned.
Technical Specifications: Gemini 4 Carbon vs. Argon
| Specification | Gemini 4 Carbon | Gemini 4 Argon |
| Developer | Google DeepMind | Google DeepMind |
| Architecture | Not disclosed | Sparse MoE |
| Parameters | Not disclosed | Not disclosed |
| Context window | Not disclosed | 1,048,576 tokens |
| Maximum output | Not disclosed | Up to 1,048,576 tokens |
| Main focus | Cybersecurity, autonomous agents | Multimodal reasoning, coding |
| Input price (introductory) | $2 / 1M tokens | $2 / 1M tokens |
| Output price (introductory) | $10 / 1M tokens | $10 / 1M tokens |
| Availability | Restricted rollout | Gemini API |
Google has not disclosed Carbon's parameter count, architecture, or context-window limits.
Cybersecurity Performance
| Benchmark | Gemini 4 Carbon |
| SEC-bench Pro — vulnerability discovery | 88.2% |
| ExploitGym — exploit validation | 54.8% |
| CWE-bench — vulnerability patching | 71.0% |
| Secure code generation | Over 2× Argon's success rate |
| Gray Swan prompt-injection success | 0.0% |
Google also reports more than a tenfold improvement in high-quality patch generation compared with an unnamed competing frontier model.
Reliability Improvements Over Argon
| Metric | Reported improvement |
| Unsupported claims | ~50% fewer |
| High-severity misalignment alerts | ~80% reduction |
| Agent overreach | 26% reduction |
| Agent under-service | 29% reduction |
| Prompt-injection success | 0.7% → 0.0% |
These results reflect Google's reported evaluations, not independently verified production performance.
Coding and Autonomous Workflows
Carbon targets multi-stage tasks involving debugging, code generation, vulnerability analysis, and automated remediation.
Key reported results:
- WebDev Arena: 1,769 Elo, fourth place.
- Web development: 89-point Elo improvement over the reported comparison baseline.
- Long-horizon execution: Multi-day autonomous engineering sessions demonstrated internally.
- Security engineering: Automated vulnerability detection and patch generation.
Availability and Pricing
The initial rollout begins during the week of October 28, 2026, through Google's Fairwind Program for selected cybersecurity organizations.
Introductory pricing is reported at $2 per million input tokens and $10 per million output tokens. Subsequent rates are expected to increase to $4 and $20, respectively. A general API release date has not been confirmed.
Conclusion
Gemini 4 Carbon improves on Argon in cybersecurity, autonomous coding, and behavioral reliability. Its reported 88.2% SEC-bench Pro score and more than twofold secure-code-generation improvement are the clearest performance indicators.
Full architectural specifications, independent benchmarks, and general availability remain pending.
