Google just dropped its latest AI model, Gemini 4 Carbon, designed to tackle complex coding jobs and cybersecurity threats. The company claims it can work through difficult engineering problems for days, finding hidden bugs and developing fixes across large codebases.
Gemini 4 Carbon also takes the top spot on Artificial Analysis’ updated Intelligence Index, scoring 59 points, pushing out Claude Opus 5.5, OpenAI’s GPT-6 Astra, and even its own predecessor Gemini 4 Argon.
What makes Gemini 4 Carbon different from Google’s previous AI model?
Carbon builds on Gemini 4 Argon with improvements in coding, reasoning, and agent-based workflows. What’s impressive is how Carbon handles long, messy coding tasks. Google says internal teams watched it work on big codebases for days straight, pushing through dead ends until it fixed the bug.
That persistence showed up on the WebDev Arena leaderboard too, where Carbon jumped 89 points. It still trails rivals on a couple of specific coding benchmarks, so it’s not perfect everywhere, but the overall trend looks strong.
Cybersecurity is its biggest selling point. Carbon scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for testing exploitability, and 71% on CWE-bench for fixing flaws.
Is Carbon safer and more reliable than Argon?
Google says it has made several changes to reduce unwanted behavior, including the tendency to invent information, take shortcuts, or act beyond a user’s instructions. Here are the most important claims.
| Improvement | What Google reports |
|---|---|
| More secure coding | More than twice the success rate for generating secure code compared with Argon in internal benchmarks. |
| Fewer unsupported claims | About half as many unsupported claims as Argon in internal testing. |
| Better task alignment | Nearly 80% fewer high-severity misalignment alerts in Google’s internal live-coding environments. |
| Better adherence to scope | Overreach fell by 26%, while under-service fell by 29%. |
| Resistance to prompt injection | A 0% attack success rate on the evaluated Gray Swan benchmark. |
When can you use Gemini 4 Carbon?
For now, Carbon is available for select cybersecurity defenders through Google’s Fairwind Program, with wider access coming later. Introductory API pricing starts at $2 per million input tokens and $10 per million output tokens. Those rates will eventually rise to $4 and $20, respectively, though Google hasn’t specified when. Its broader availability will depend on how Google evaluates its safeguards and real-world performance.
RELATED COVERAGE
