Gemini 4 Carbon Sets Cybersecurity Benchmark Records
Google DeepMind's Gemini 4 Carbon leads the Artificial Analysis Intelligence Index and sets records across three cybersecurity benchmarks.
👁 124 reads
Google DeepMind announced Gemini 4 Carbon on October 28, 2026. The company calls it its most capable model, and it took first place in the Artificial Analysis Intelligence Index v5.0 with a score of 59.
Carbon arrived just two weeks after Gemini 4 Argon, showing how quickly DeepMind is pushing the frontier. The model focuses on two areas: long-running engineering tasks that require days of agentic work, and cybersecurity.
Three cybersecurity benchmarks, three records
In cybersecurity, Carbon achieved state-of-the-art results across key stages of defense, from finding vulnerabilities to validating and fixing them. It scored 88.2% on SEC-bench Pro (vulnerability detection), 54.8% on ExploitGym (exploitation) and 71.0% on CWE-bench (bug fixing). In internal tests, Carbon generates secure code more than twice as effectively as Argon. The Chrome team also said Carbon performed more than 10 times better than a leading model from another lab when generating high-quality patches.
Coding and agentic tasks
Carbon gained 89 Elo points in programming and ranked fourth in WebDev Arena with a score of 1769. It is designed for complex, multistep engineering tasks. DeepMind’s internal teams have seen Carbon work for entire days, pushing through dead ends to solve problems such as finding and fixing subtle bugs in large codebases.
Gradual rollout and pricing
DeepMind is taking a cautious approach to releasing the model. Carbon is part of the US government’s voluntary pre-release testing process, and initial access is going to selected cyberdefense specialists through the Fairwind program. After gathering feedback and refining safeguards, DeepMind plans to make the model available to developers, businesses and consumers.
Pricing is the same as Argon’s: $2 per million input tokens and $10 per million output tokens. Cached input tokens cost 95% less than the standard input price.
DeepMind’s focus on cybersecurity, combined with a slow, controlled rollout through Fairwind, comes as regulators increasingly ask who is responsible for models capable of finding and exploiting vulnerabilities in systems.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →More about Gemini 4 Carbon
For a full overview of pricing, access and the model, see the Gemini 4 Carbon hub. I cover all 19 benchmarks and explain their results in the Gemini 4 Carbon benchmark analysis, and compare the model with its predecessor in Gemini 4 Carbon vs. Argon.
More on this from QUANTSLANT
benchmarks·Gemini 4 Carbon benchmarks: all 19 tests and where it losesGoogle reports Gemini 4 Carbon leads in 17 of 19 benchmarks. Here’s the full score table, what each test measures and where rivals win.