Google is moving quickly. Gemini 4 Argon was released just two weeks ago, and then Google came out with another update of its product – Gemini 4 Carbon on 28 October. This one promises to be the strongest version so far, but this time, it is focused on cybersecurity, engineering, and safety behaviors. Also, there will be less fabrication involved in the process. The product is topping some independent index now. However, the launch of the new version is much more limited compared to that of Argon. This is what changed between the two Gemini 4 versions.
Also read: Generative AR explained: The next step after AI image generators
Among the two, carbon seems to be the one that is more versatile and more specialized. In the definition by Google, it is described as their best version due to having advanced cybersecurity and long-horizon capabilities. It is also ranked number one in the Artificial Analysis Intelligence Index v5.0 with a score of 59. The new version is built upon Argon.
When measured internally by Google, Carbon is over twice as efficient as Argon in creating secure code. When tested externally, but without any guardrails, its performance is 88.2 percent on SEC-bench Pro for vulnerability discovery, 54.8 percent on ExploitGym for exploitation, and 71.0 percent on CWE-bench for patching. In addition, according to Google, the Chrome team identified that Carbon outperformed another lab’s massive frontier model in the creation of quality patches by more than 10 times. Wiz has reported that Carbon is superior on web penetration testing benchmarks too.
Carbon is designed for long-term agential workflow. According to Google, it was observed by internal teams for several days running until it found and fixed the subtle bugs within the big codebase and the low-level systems. Moreover, it performs better at web development, rising 89 Elo points to become fourth in WebDev Arena with 1,768 points.
Also read: GitHub Octoverse 2026: India is the second largest developer hub and growing faster than the US
The other area in which Google has concentrated on is alignment. Argon had previously topped the Gray Swan indirect prompt injection benchmark with an attack success rate of 0.7 per cent. The score for Carbon on this test was 0.0 per cent. For the Terminal-Bench 4.0 test, the cheating rate became 0.0 per cent, while the attempt rate reduced to 0.4 per cent.
Additionally, according to Google, the claims made by Carbon are only half those made by Argon, overreach reduced by 26 per cent and under-service by 29 per cent.
However, the price remains the same, at least for now. The price of Carbon will be launched at the introductory price offered by Argon: $2 for every million of input tokens and $10 for every million of output tokens. Cached input tokens will be available at 95% cheaper rates. After the introductory period, the price will go up to $4 and $20, respectively.
The issue of access is the most notable difference between the two. For this week, Carbon will be made available only to a select few cyber security defenders with gradual inclusion of trusted partners using the Google Fairwind Program. Google is also taking part in the pre-release testing program by the US government on a voluntary basis. Carbon will be made accessible to developers, businesses and consumers at the earliest possible opportunity, without any specified date.
For anyone building right now, Argon is the one you can actually use. Carbon is better read as a marker of Google’s direction: security-first, longer-running agents and a more cautious release strategy. Wider availability will show whether its benchmark lead holds up in everyday work.
Also read: US vs China AI race: What the Chinese stock crash really changes