Gemini 4 Carbon: what’s new and what people are building.
Google’s Gemini 4 Carbon, announced 28 October 2026: what’s new for coding, agents and security, who can use it today, and the first builds shared on X.
By The Model Atlas · 5 sources
In this post5 sections
Google announced Gemini 4 Carbon on 28 October 2026, calling it its next frontier model and its strongest yet, with a particular focus on cybersecurity and long-horizon work. It arrives two weeks after Gemini 4 Argon, and it is not open to everyone yet. Here is what Google has confirmed, who can use Carbon today, and what builders have already made with it.
What is Gemini 4 Carbon?
Gemini 4 Carbon is a frontier AI model from Google DeepMind, announced on Google’s blog by Koray Kavukcuoglu, SVP of Google DeepMind and Google’s Chief AI Architect. Google describes it as building on Gemini 4 Argon, and as its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities.
Google’s announcement puts two things first. Carbon is built for sustained agentic work: long, multi-step engineering tasks rather than single answers. And it goes first to cyber defenders, the people who find and fix vulnerabilities in real software, before a wider release.
What’s new in Gemini 4 Carbon
Long-horizon coding and agents
Google says Carbon was designed for much harder, long-horizon engineering problems. Its own teams have seen it run for days, working past dead ends on multi-step tasks such as root-causing and patching subtle bugs across large codebases and low-level systems. Google also lists improvements in agentic and scientific tasks, full-stack web development and multimodal understanding.
For web development specifically, Google reports that Carbon gained 89 Elo points on WebDev Arena, placing it fourth with a score of 1768. On version 5.0 of the Artificial Analysis Intelligence Index, Google says Carbon ranks first with a score of 59.
Cybersecurity, from finding bugs to fixing them
Most of the announcement is about security. Google frames Carbon’s results around the full defence lifecycle and reports these external benchmark scores, measured without guardrails:
- Finding vulnerabilities: 88.2% on SEC-bench Pro, which Google calls state of the art.
- Validating them through exploitation: 54.8% on ExploitGym, also a new state of the art according to Google.
- Patching the root cause: 71.0% on Collinear’s CWE-bench.
On internal benchmarks, Google says Carbon more than doubles Argon’s success rate at generating secure code. It adds that the Chrome team reported Carbon performing more than ten times better than another lab’s large frontier model at producing high-quality patches, and that Wiz sees Carbon leading on web penetration-testing benchmarks.
Safety and honesty
Carbon uses the same layered safety stack as Argon: model training, real-time misuse defences, prompt-injection resilience and chain-of-thought monitoring. Google reports several behaviour changes on top of that:
- Successful cheating on Terminal-Bench 4.0 fell to 0.0%, with cheating attempts at 0.4%, while overall accuracy improved.
- In live coding with Google’s internal teams, high-severity misalignment alerts dropped by almost 80%.
- Unsupported claims were cut by roughly half compared with Argon in internal testing.
- Overreach fell by 26% and under-service by 29%, so multi-step plans stay within scope more reliably.
- On Gray Swan’s Indirect Prompt Injection benchmark, the attack success rate was 0.0%, down from Argon’s 0.7%.
What people are building with Gemini 4 Carbon
Even with limited access, builders have started sharing what Carbon made for them. The Atlas files every build under the model its creator credits, and the Gemini 4 Carbon page collects them all. So far they lean heavily towards 3D, much of it in three.js:
- A city-scale 3D model of Oxford in three.js, which Shuyang (Kevin) Sun says Carbon generated without any human intervention.
- A robotic eye that blinks, focuses and then explodes into its parts, built by noclipepe from one prompt in about 30 minutes.
- A detailed 3D scene made with Blender MCP and three.js by Chouaieb Nemri.
- A cozy virtual pet game themed around Cambrian sea creatures by Michael Joffe.
- A 3D helicopter and a drivable seaplane on water from Lumina’s Arena testing.
- Two black hole horizons and a cloud-like cat chasing a butterfly, both made fully in three.js by Bee during Arena testing.
- A side-by-side 3D test against Claude Opus 5.5 by Salio.
Each build page shows the creator’s original post on X, with their name and a link back. Several of these came from testing on Arena, the model comparison site; Google’s announcement does not mention Arena, so availability there can change at any time.
How to try Gemini 4 Carbon
Google is taking what it calls a deliberate approach to rollout. It is taking part in the US government’s voluntary process for pre-release model testing, and access is opening in stages:
- Cyber defenders first. Starting the week of the announcement, Google is making Carbon available to a subset of cyber defenders.
- Trusted partners next. Access then widens gradually to more partners through Google DeepMind’s Fairwind Program (opens in a new tab), which gives high-priority defenders such as governments, healthcare providers and telecommunications services early access to advanced models. Organisations can apply on that page; Google vets every applicant.
- Everyone else later. Google says it will bring Carbon to developers, businesses and consumers through its products as soon as possible, after gathering more feedback on guardrails. It has not given a date.
For pricing, Google says Carbon will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper than regular input. After the introductory period, the price becomes $4 per million input tokens and $20 per million output tokens.
Until then, the closest generally available model is Gemini 4 Argon, which Google DeepMind lists as available to Google AI Pro and Ultra subscribers. You can see what people have built with it on the Gemini 4 Argon page, and find free ways into Google’s apps in the Google Gemini and Google AI Studio directory entries.
Gemini 4 Carbon FAQ
When was Gemini 4 Carbon released?
Google announced Gemini 4 Carbon on 28 October 2026. Google said it would start making it available to a subset of cyber defenders that week; wider availability has no date yet.
Can I use Gemini 4 Carbon in the Gemini app or the API?
Not yet. Google says it will bring Carbon to developers, businesses and consumers through its product surfaces as soon as possible, once it has gathered more feedback on guardrails.
How much will Gemini 4 Carbon cost?
Google says the introductory price will match Argon: $2 per million input tokens and $10 per million output tokens, with cached input 95% off. After the introductory period it rises to $4 and $20 per million input and output tokens.
What is Gemini 4 Carbon best at?
Google highlights long-horizon coding and agentic tasks, and cybersecurity defence: finding, validating and patching vulnerabilities. Early builds shared on X are mostly 3D scenes and small games, many written in three.js.
How is Carbon different from Gemini 4 Argon?
Google positions Carbon as the stronger model, with gains in coding, agents and security, more than twice Argon’s success rate at generating secure code on internal benchmarks, and fewer unsupported claims. Argon is the one you can use today.
We’ll update this post when Google opens Carbon to more people. In the meantime, new builds appear on the Gemini 4 Carbon page as creators share them.
Sources · 5
- Gemini 4 Carbon: our next frontier model (Google, The Keyword, 28 October 2026) (opens in a new tab)
- Fairwind Program (Google DeepMind) (opens in a new tab)
- Gemini models (Google DeepMind) (opens in a new tab)
- Gemini 4 Argon: our next era of frontier intelligence (Google, The Keyword) (opens in a new tab)
- Gemini 4 Safety and Security report (Google DeepMind, PDF) (opens in a new tab)