Google just entered the frontier model race with Gemini 4 Carbon, released October 29, 2026. It follows Gemini 4 Argon by two weeks and, according to Google’s own benchmarks, lands ahead of Claude Opus 5.5 and GPT-6 Astra on several key measures.
Benchmark Claims
Google reports Carbon hit an AA Index v5.0 score of 59, the highest of any model on that general AI benchmark. The company also claims a near sweep across agentic coding, computer use, knowledge work, science and math, and several other categories.
Cybersecurity is where Carbon pulls furthest from the competition. Google says the model excels at finding and prioritizing vulnerabilities and patching them at the root cause. It’s also twice as likely to generate secure code compared to Argon, according to Google’s figures.
Safety and Security Changes
Beyond raw performance, Google highlights several safety improvements in Carbon:
- Reduced shortcuts and loophole exploitation
- Improved transparency and honesty, including scoring on an unsupported claims test
- Stronger resistance to prompt injection
- Better collaboration behavior, described as following through without overstepping
Rollout Plan
Carbon will not be widely available immediately. Google is following the same staged rollout it used with Argon: initial access goes to select cybersecurity defense companies for pre-release testing, then broader access through the company’s Fairwind Program, then a general release. Argon completed that same path and is now live in the Gemini app and AI search mode.
If you’re tracking the frontier model arms race, Carbon is the current leader on Google’s scoreboard. Whether those benchmarks hold up in real operator workflows is a different question.
