Google has announced Gemini 4 Carbon, a new artificial intelligence model. The company calls it its strongest model so far.
The news came on Oct. 28 in a blog post by Koray Kavukcuoglu. He is a senior vice president at Google DeepMind and Google’s chief AI architect.
Carbon follows Gemini 4 Argon, which Google launched two weeks earlier. Google says the new model shows gains in coding, long tasks and cybersecurity.
Benchmark Results for Gemini 4 Carbon
Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. The index compares AI models across a range of tasks.
The model also scored well on security tests. It reached 88.2% on SEC-bench Pro, which measures how well a model finds software flaws.
It scored 54.8% on ExploitGym, a test of whether flaws can be exploited. It also reached 71.0% on CWE-bench, which checks how well a model fixes the root cause of a flaw.
Google said Carbon more than doubled Argon’s success rate at writing secure code on internal tests. The Chrome team reported Carbon did more than 10 times better than another lab’s large model at writing quality patches.
In web development, Carbon gained 89 elo points. It now holds fourth place on WebDev Arena with a score of 1768.
Google said its teams have seen Carbon run for days on hard engineering tasks. These include finding and fixing hidden bugs across large codebases.
Safety Features and Rollout Plans
Google said Carbon uses the same safety system as Argon. That system includes model training, real-time misuse defenses and monitoring of the model’s reasoning.
On Terminal-Bench 4.0, successful cheating by the model fell to 0.0%. Attempts to cheat dropped to 0.4%.
In internal coding work, high-severity misalignment alerts fell by almost 80%. Google also said Carbon makes about half as many unsupported claims as Argon.
Carbon cut overreach by 26% and under-service by 29%, according to the company. On Gray Swan’s prompt injection test, it blocked every attack attempt, compared with Argon’s 0.7% attack success rate.
Pricing will match Argon’s introductory rate. That is $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off.
Google is taking part in the U.S. government’s voluntary process for testing AI models before release.
Starting this week, Carbon will be available to a small group of cyber defenders. Google will then slowly add trusted partners through its Fairwind Program.
The company said it will keep gathering feedback and improving guardrails. It plans to release Carbon to developers, businesses and consumers as soon as possible.
