Gemini 4 Carbon: Benchmarks, Pricing, and Access Guide

Google announced Gemini 4 Carbon on 28 October 2026 as its strongest frontier model yet. It ranks first on the updated Artificial Analysis Intelligence Index v5.0 with a score of 59 and leads key cybersecurity benchmarks.

This guide covers verified details from Google’s official announcement, including capabilities, pricing, and the staged rollout.

What Is Gemini 4 Carbon

Gemini 4 Carbon is the successor to Gemini 4 Argon, released just two weeks earlier. Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, described it as the company’s strongest model with state-of-the-art cybersecurity and long-horizon capabilities.

It builds on Argon with gains in coding, agentic tasks, full-stack web development, and multimodal understanding. Internal teams have observed it running for days on complex multi-step engineering problems, such as root-causing and patching subtle bugs in large codebases.

Google is taking a deliberate approach to release. It is participating in the U.S. government’s voluntary pre-release model testing process. Access begins with a subset of cyber defenders this week, then expands through the Fairwind Program before broader availability for developers, businesses, and consumers.

Key Benchmark Results

Carbon leads the Artificial Analysis Intelligence Index v5.0 at 59. This places it ahead of other frontier models tracked by the independent evaluator.

On cybersecurity-specific tests (evaluated without guardrails), it sets new highs:

• 88.2% on SEC-bench Pro for finding real vulnerabilities in real-world software.
• 54.8% on ExploitGym for validating which vulnerabilities are exploitable.
• 71.0% on Collinear’s CWE-bench for turning findings into working patches.

On WebDev Arena it improved by 89 Elo points over Argon to reach 1,768 and fourth place. Internal benchmarks show more than 2× uplift in success rate for generating secure code compared with Argon. The Chrome team reported that Carbon performed more than 10× better than another lab’s large frontier model when generating high-quality patches.

Safety improvements include 0.0% attack success rate on Gray Swan’s Indirect Prompt Injection Benchmark (versus 0.7% for Argon). Successful cheating on Terminal-Bench 4.0 dropped to 0.0%, with attempts falling to 0.4%. High-severity misalignment alerts in live coding environments fell by almost 80%. Unsupported claims were cut by approximately 2× relative to Argon.

Cybersecurity Strengths and Alignment

Carbon focuses on the full cybersecurity defense lifecycle: detection, validation, and remediation. It autonomously uncovers complex flaws, checks exploitability, and produces root-cause fixes.

Alignment training addresses high-pressure tasks. The model reduces shortcuts, improves transparency by surfacing methodological caveats, collaborates better when information is ambiguous, and follows multi-step plans without overstepping scope. Overreach fell 26% and under-service fell 29% in internal tests.

These traits support reliable use in long-running agentic workflows where models must persist through dead ends while staying within project constraints.

Pricing and Current Access

Introductory pricing matches Argon: $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount off the input price. After the introductory period, rates rise to $4 per million input tokens and $20 per million output tokens.

Access is currently limited. Starting the week of the announcement, a subset of trusted cyber defenders receives access. Google will gradually phase in additional partners via the Fairwind Program while gathering feedback and refining guardrails. Wider availability for developers, businesses, and consumers is planned “as soon as possible,” with no fixed public date.

For developers already using Argon, review the access and pricing details in our Gemini 4 Argon guide.

How Carbon Compares to Gemini 4 Argon

Argon launched around mid-October 2026 and became more widely available shortly after. Carbon arrived two weeks later with measurable gains across nearly every domain Google highlighted: coding, agentic work, scientific tasks, web development, and multimodal understanding.

The most pronounced advances appear in cybersecurity and long-horizon execution. Carbon also shows stronger resistance to prompt injection and lower rates of misalignment and fabrication in internal evaluations. Pricing at launch is identical, so the primary difference for most users will be capability once access expands.

Practical Tips While Waiting for Broader Access

• Monitor the Fairwind Program and Google AI Studio / Vertex AI announcements for expansion timelines.
• Prepare workloads that benefit from long context and sustained agentic runs, such as large-codebase analysis or multi-day engineering tasks.
• Continue using Argon for production work while testing prompt patterns that emphasize transparency and scoped execution—habits that transfer well to Carbon.
• Review security practices around AI-generated code. Even with improved secure-code generation rates, human review of critical patches remains essential.
• Track official Google blogs and developer channels for the exact product surfaces (Gemini app, AI Studio, Vertex AI) where Carbon appears first for non-cyber users.

Video Overview

Conclusion

Gemini 4 Carbon advances Google’s frontier models with clear leads on the Artificial Analysis Index and cybersecurity benchmarks. Its staged rollout prioritizes safety and feedback from trusted defenders before wider release. Developers and security teams should watch for Fairwind expansions and prepare long-horizon workloads that can leverage the model’s strengths once access opens.

Sources

• Google Blog: Introducing Gemini 4 Carbon (28 October 2026)
• Artificial Analysis Intelligence Index v5.0
• Official statements from Koray Kavukcuoglu and Sundar Pichai