Google Gemini 4 Carbon Tops Benchmarks, Cyber Defenders Get It First
3 min readGoogle DeepMind introduced Gemini 4 Carbon on October 28, calling it the company’s strongest model yet and the new state of the art for cybersecurity and long-running engineering work. Carbon arrives just two weeks after Gemini 4 Argon. The twist is who gets it: for now, access is limited to a vetted group of cyber defenders, not the public.
What Came Before Gemini 4 Carbon
Argon launched in mid October as Google’s flagship Gemini 4 release, with a multi-layered safety stack covering training, misuse defenses, prompt-injection resistance and chain-of-thought monitoring. Carbon builds on that foundation and shifts the emphasis toward finding, exploiting and patching software flaws, the kind of work that can cut both ways depending on who holds the model.
What Happened
Google reports 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. Internal tests showed more than double Argon’s success rate at writing secure code, and the Chrome team said Carbon produced high-quality patches more than ten times as often as a rival lab’s frontier model.
Independent evaluators agree it is a frontier model. Artificial Analysis ranks Carbon first on its new Intelligence Index v5 with a score of 59, leading four of nine evaluations, and it holds the top spot on the Arena text leaderboard. Its coding agent score of 59 still trails Claude Code with Opus 5.5 at 62, and its hallucination rate rose to 11 percent from Argon’s 4 percent, according to Latent Space.
On safety, Google says cheating on Terminal-Bench 4.0 fell to zero, high-severity misalignment alerts in internal coding dropped by almost 80 percent, and prompt-injection attacks succeeded 0.0 percent of the time on Gray Swan’s benchmark. Google is also taking part in the United States government’s voluntary pre-release testing process.
Why It Matters
Access is the story. Carbon goes first to a subset of cyber defenders through Google’s Fairwind Program, then gradually to more trusted partners. Developers, businesses and consumers come later, with no date given. Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the launch period.
The staged rollout is becoming the norm for the most capable frontier models and signals that labs now treat offensive security skill as a reason to gate a release. Watch for how quickly Fairwind expands and whether the higher hallucination rate holds up under outside testing.
Google has set the pace on cyber benchmarks for the moment, and the people who get to use that lead are being chosen carefully.