Google DeepMind has introduced Gemini 4 Carbon, its strongest model to date, designed for long-horizon agentic coding and cybersecurity defense. The model ranks #1 on the updated v5.0 Artificial Analysis Intelligence Index with a score of 59.
Key points
- Gemini 4 Carbon ranks #1 on the AI Analysis Intelligence Index.
- The model excels in cybersecurity defense and agentic coding tasks.
- Carbon outperforms its predecessor Gemini 4 Argon on most benchmarks.
- Access to Gemini 4 Carbon is currently restricted to select cyber defenders.
- Google DeepMind developed Gemini 4 Carbon for long-running engineering work.
Access to Gemini 4 Carbon is currently limited to a vetted group of cyber defenders, with no public API availability. Parameter count and context window details have not been disclosed. Unlike its predecessor, Gemini 4 Argon, which has a 1,048,576-token context, Carbon's specifics remain unannounced.
Gemini 4 Carbon runs exclusively on Google-hosted infrastructure, with no open weights, self-hosting options, or quantized variants available. It outperforms its predecessor, Gemini 4 Argon, on 15 out of 17 published benchmarks, including GPT-6 Astra, Opus 5.5, and Fable 5.1.
By the numbers
- AI Analysis Intelligence Index score
- 59
- BFS F1 on GraphWalks
- 96.2%
- benchmarks led
- 15 of 17
The model achieved a 96.2% BFS F1 score on GraphWalks (256K to 1M tokens), significantly higher than Argon's 84.2%. However, it scored 81.5% on OSWorld-2.1, trailing Claude Opus 5.5 at 87.2%. Its key strengths lie in state-of-the-art vulnerability finding, exploitation, and patching.
Gemini 4 Carbon is the second model in the Gemini 4 series, following Gemini 4 Argon. Google states the model is engineered for challenging, long-duration tasks, with internal teams observing it running for days on single tasks. These tasks include root-causing and patching subtle bugs in large codebases and low-level systems.
Google has also tuned Carbon for full-stack web development and multimodal understanding. On the WebDev Arena, Carbon gained 89 Elo points, reaching #4 with a score of 1769. On Terminal-Bench 4.0 (AA), Carbon scored 65.2% in agentic coding, surpassing Opus 5.5's 59.6%. It achieved 9.6% on SWE-sweep, narrowly ahead of GPT-6 Astra's 9.5%, though GPT-6 Astra leads on FrontierSWE at 65.5% compared to Carbon's 61.2%.
Sources and references
Original report
Google DeepMind Introduces Gemini 4 Carbon, Its Top Model for Cyber Defense and Coding
MarkTechPost
Read on MarkTechPost (opens in a new tab)Linked in the original report
- Linkblog.google/innovation-and-ai/models-and-research/g… (opens in a new tab)
- Linkartificialanalysis.ai (opens in a new tab)
- Linkai.google.dev/gemini-api/docs/models/gemini-4-argon (opens in a new tab)
- Linkblog.google/innovation-and-ai/models-and-research/g… (opens in a new tab)
- Linkarena.ai/leaderboard/code/webdev (opens in a new tab)
- Paperstorage.googleapis.com/deepmind-media/gemini/gemini_4_carbon_m… (opens in a new tab)
Ref 00EMF495
Cite this
APA
Interestana AI Editorial. (2026, October 28). Google DeepMind Unveils Gemini 4 Carbon for Cyber Defense. Interestana. https://interestana.com/articles/google-deepmind-introduces-gemini-4-carbon-its-top-model-for-cyber-defense-and-coding-00emf495
MLA
“Google DeepMind Unveils Gemini 4 Carbon for Cyber Defense.” Interestana, 28 Oct. 2026, interestana.com/articles/google-deepmind-introduces-gemini-4-carbon-its-top-model-for-cyber-defense-and-coding-00emf495.
Chicago
Interestana AI Editorial. “Google DeepMind Unveils Gemini 4 Carbon for Cyber Defense.” Interestana, October 28, 2026. https://interestana.com/articles/google-deepmind-introduces-gemini-4-carbon-its-top-model-for-cyber-defense-and-coding-00emf495.
BibTeX
@misc{interestana_00emf495,
author = {{Interestana AI Editorial}},
title = {{Google DeepMind Unveils Gemini 4 Carbon for Cyber Defense}},
howpublished = {Interestana},
year = {2026},
month = oct,
url = {https://interestana.com/articles/google-deepmind-introduces-gemini-4-carbon-its-top-model-for-cyber-defense-and-coding-00emf495},
note = {Summary of reporting by MarkTechPost, https://www.marktechpost.com/2026/10/28/google-deepmind-gemini-4-carbon-frontier-model/}
}Credit the source
Original reporting: MarkTechPost, “Google DeepMind Introduces Gemini 4 Carbon, Its Top Model for Cyber Defense and Coding”, https://www.marktechpost.com/2026/10/28/google-deepmind-gemini-4-carbon-frontier-model/(opens in a new tab)
Follow this story
Get Artificial Intelligence in your inbox — new stories in 40 words or fewer, once a week.
Subscribe