Google Gemini 4 Carbon Guide

Gemini 4 Carbon is Google’s strongest AI model, made available on November 10, 2026. Google built it for the hardest work: long-running agent tasks in software engineering, knowledge work, and science, and for cybersecurity defence.

Argon for most work. Carbon when the problem is hard enough to need it.

Gemini 4 Argon

The better fit for most coding and knowledge work

Gemini 4 Carbon

For harder tasks such as complex multi-step debugging and open-ended research; it uses more tokens to explore, reason, and check its work

Gemini app and AI Mode in Search

Google AI Ultra subscribers

Google Antigravity

Ultra subscribers and enterprises, with a /swarm preview that launches a team of agents from one prompt

Developers

Gemini API paid tier in Google AI Studio and Android Studio

Enterprises

Gemini for business and CodeMender

Until November 30, 2026

$2 per million input tokens and $10 per million output tokens, the same as Argon

From November 30, 2026

$4 per million input tokens and $20 per million output tokens

Carbon also uses more tokens per task than Argon, so real costs can be higher than the price list suggests.

• First on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, at lower cost per task than the next-best outside model
• Top results on two financial research evaluations
• Leading scores on security benchmarks for finding, exploiting, and patching vulnerabilities

These are Google’s own figures. Test Carbon on your own tasks before you rely on them.

• Carbon first went to cybersecurity defenders through Google’s Fairwind Program about two weeks before launch
• Defenders can apply to Fairwind for cyber access
• Scientists can apply for six months of free access to Argon and Carbon in Antigravity’s highest usage tier

• Google says Carbon blocked 99% of attack attempts in simulated cyberattack tests
• It passed 92% of tasks on an agent honesty benchmark, versus 70% for Argon
• Google says resistance to prompt injection still needs improvement

• Give the whole goal, the context, and what done looks like
• Ask for a plan, then let it work in stages
• Tell it how to verify results, such as running tests
• Keep approvals on for risky actions

Investigate why checkout fails for some users after yesterday’s release. Reproduce the bug, find the root cause across the three services, propose a fix with tests, and summarise what you ruled out.

• Everyday coding, writing, and analysis
• Tasks where speed and cost matter more than depth
• Anything Argon already does well