Google has just announced what is potentially the most powerful AI model in the world—at least according to early metrics. The new model, named Gemini 4 Carbon, comes just two weeks after the release of Gemini 4 Argon, signaling that Google is aggressively matching the rapid release pace previously dominated by competitors like OpenAI and Anthropic.
Setting a New State-of-the-Art Standard
Built directly on the foundation of Gemini 4 Argon, Carbon is designed to be a frontier model with two major areas of focus:
1. “Long Horizon Work”
One of the most significant frustrations with current AI agents is their tendency to stall, quit, or hallucinate when a task becomes too complex. Google claims Carbon excels in “long horizon work,” meaning it can handle massive, complicated tasks that require multiple steps and extended periods to finish. According to Google’s internal teams, Carbon has been observed running for days on complex engineering tasks, actively pushing through dead ends rather than giving up. This level of endurance could fundamentally change how much work developers can hand off to autonomous agents.
2. Advanced Cyber Security
Cyber security is clearly Google’s primary focus with Carbon. The model is specifically tuned to find security holes in software, confirm their validity, and patch them. It swept all three of Google’s internal security tests covering vulnerability detection, exploitation, and patching.
Benchmark Dominance (and a Few Misses)
Carbon is putting up staggering numbers on industry benchmarks. Out of 18 benchmarks shared by Google, Carbon beat heavyweights like GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 on 16 of them. Furthermore, it outperformed its predecessor, Argon, across every single metric.
The only areas where Carbon did not claim the absolute top spot were:
- Coding: Trailed slightly behind GPT-6 Astra.
- Computer Use Test: Trailed Claude Opus 5.5.
- Web Development: Placed 4th on the WebDev Arena.
However, its most impressive flex comes from security and integrity tests. On Grey Swan’s prompt injection benchmark—which tests if bad actors can trick the model with hidden instructions—Carbon scored a 0% attack success rate. In another test measuring how often an AI will “cheat” to fake a passing result on coding work, Carbon again scored a perfect 0%. Google also reports that Carbon makes about half as many unsupported claims as Argon, preferring to admit missing information rather than hallucinating evidence.
Release Timeline and Pricing
Currently, Carbon is not available to the general public. In standard industry practice for powerful frontier models, Google is first rolling out Carbon to a small group of cyber defenders. This allows “good guys” to patch critical software vulnerabilities before the model falls into the hands of potential bad actors who could exploit those same flaws.
While there is no official public release date yet, if Google follows the same playbook they used for Argon (which went to defenders on Sept 30th and the public on Oct 13th), a broader release could happen in roughly two weeks.
When it does launch for developers, Carbon will initially share Argon’s introductory pricing: $2 per million input tokens and $10 per million output tokens. After the introductory period, the price is set to double.
The Naming Confusion
The one lingering question is Google’s naming convention. With Gemini 4 Argon and Gemini 4 Carbon both released within weeks of each other, it is unclear if Google is moving toward a tiered naming system (similar to OpenAI) or simply adopting new elemental names for every iterative release.
Regardless of the naming structure, Gemini 4 Carbon appears to be an incredibly powerful agentic AI that refuses to quit. If it lives up to its on-paper promises, it will be a massive asset for digital security and long-term automation.


Leave a Reply