Gemini 4 Carbon: Google's strongest model, built for cyber defense
Google DeepMind announced Gemini 4 Carbon on October 28, 2026, and calls it the company's strongest model yet. It is built around two capabilities: long-horizon agentic work and cybersecurity defense. Carbon builds on Gemini 4 Argon, which shipped two weeks earlier, and DeepMind says it lands at number one on the recently updated v5.0 of the Artificial Analysis Intelligence Index, with a score of 59 and strong gains across coding and agentic tasks.
The rollout is deliberately slow. Google says it is taking part in the U.S. government's voluntary process for pre-release model testing, that Carbon starts this week with a subset of cyber defenders, and that wider access follows gradually through its Fairwind Program, which targets trusted partners. General availability for developers, businesses, and consumers is promised, without a date.
Cyber benchmarks, end to end
The benchmark story is the center of the announcement. DeepMind says Carbon's biggest jump is across the cybersecurity defense lifecycle: finding vulnerabilities, validating whether they are exploitable, and patching the root cause. The reported external numbers were evaluated without guardrails:
- SEC-bench Pro: 88.2%, which DeepMind calls state of the art for autonomously uncovering vulnerabilities in real-world software
- ExploitGym: 54.8%, a new lead on a popular exploitation benchmark that defenders use to prioritize which flaws to fix first
- CWE-bench, from Collinear: 71.0%, turning validated findings into working code fixes
Internal numbers follow the same pattern. Google reports that Carbon achieves more than 2x the success rate of Argon on generating secure code, on its own benchmarks. The Chrome team reported that Carbon produced high-quality patches that were more than 10x better than a large frontier model from another lab, per the company. Wiz, Google's cloud security platform, also observes Carbon leading on web penetration testing benchmarks, a core capability of its Scan for Good initiative. Those internal results come without published eval details, so they read as company numbers until someone else reproduces them.
Built to work for days
Carbon targets sustained agentic workflows. Google says its own teams have watched it run for days, pushing through dead ends to work on multi-step engineering tasks like root-causing and patching subtle bugs across large codebases and low-level systems. On WebDev Arena, the model jumps 89 elo points to land fourth with a rating of 1768, according to DeepMind.
A heavier safety layer
Carbon rides on the same multi-layered stack as Argon: training-time safety, real-time misuse defenses, prompt injection resilience, and chain-of-thought monitoring. Google's claimed deltas for Carbon:
- On Terminal-Bench 4.0, successful cheating dropped to 0.0% and cheating attempts to 0.4%, while overall accuracy improved
- High-severity misalignment alerts in internal live coding environments fell by almost 80%
- Unsupported claims dropped roughly 2x versus Argon. Carbon is meant to put methodological caveats into the deliverable instead of fabricating or hedging around missing evidence
- Overreach fell 26% and under-service 29%, which Google frames as more reliable execution of multi-step plans
- On the Gray Swan Indirect Prompt Injection Benchmark, 0.0% attack success, versus 0.7% for Argon
Like the benchmarks above, these are self-reported. Google published a safety report for Argon and positions Carbon as an alignment step further on top of that work.
Pricing, and what it means
Carbon launches at the same introductory price as Argon: $2 per million input tokens and $10 per million output, with cached input tokens at 95% off the input price. After the introductory period, pricing doubles to $4 and $20 per million tokens.
Two things to take from this. First, if you route workloads through an API, Carbon is not in the pool yet, and the useful signal is the price anchor: the frontier tier of the Gemini 4 family settles at $2/$10 introductory. Second, if you run a security team, the Fairwind Program is the door to watch. It is the path to a model that Google claims leads on vulnerability finding, exploitation, and patching, in the exact order a defense pipeline needs them.
Source: deepmind.google
More news
Have something in mind?
Start a conversation