From the source
Google debuted Gemini 4 Carbon, a frontier large language model built for complex programming and cybersecurity work, two weeks after its previous flagship, Gemini 4 Argon.
- Google says internal tests show Carbon is more than twice as good at generating secure code, and that it outperforms GPT-6 Astra and Claude Opus 5.5 across multiple benchmarks.
- Its most notable result is a record score on the Artificial Analysis Intelligence Index, tested on a new version of the index released the same day that adds two benchmarks and drops saturated questions.
- Carbon also beat the previous record on SEC-bench Pro, which measures how well models find weak points in complex software, and is better than Argon at patching what it finds.
- Koray Kavukcuoglu,
Google DeepMind's senior vice president, wrote that the Chrome team reported Carbon performed more than 10X better than another lab's large frontier model when generating high quality patches.
- Google says internal teams using Carbon in live coding environments reported a nearly 80% drop in misalignment alerts, and that it cuts unsupported claims by roughly 2x over Argon.
- On Terminal-Bench 4.0 the model tried to cheat on only 0.4% of questions, none successfully, and it blocked all simulated prompt injections in the Gray Swan IPI Benchmark.
- Google will make Carbon available later this week through its Fairwind Program, which gives organizations such as internet providers early access, and plans to refine its guardrails with early adopter feedback before broad availability.