Google introduced Gemini 4 Carbon on October 28, claiming the top spot on a leading AI intelligence index and its strongest gains in cybersecurity. The catch: most people cannot use it yet.
Carbon scored 59 on Artificial Analysis Intelligence Index v5.0, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. Google says a small group of cyber defenders will receive access first, with more trusted partners to follow through its Fairwind Program.
More than a leaderboard bump
Carbon follows Gemini 4 Argon just two weeks after its launch. Beyond the overall ranking, Google reports improvements in long-running engineering work, computer use, science and multimodal tasks. In its benchmark table, Carbon scores 65.2% on Terminal-Bench 4.0 against Argon’s 57.1%, and 96.2% on the long-context GraphWalks test against 84.2%.
Overall model comparison, based on scores in Google’s October 28 benchmark release.
Google also says its teams have run Carbon for days on difficult software engineering tasks, including tracking down subtle bugs and building patches across large codebases. Those internal observations are promising, but they are not a substitute for independent testing in production.
The cybersecurity numbers stand out
On SEC-bench Pro, which tests vulnerability discovery, Carbon reached 88.2%, compared with 85.4% for GPT-6 Astra and 83.4% for Argon in Google’s published results.
Vulnerability discovery benchmark. Source: Google DeepMind.
On ExploitGym, Carbon posted 54.8% with a ten-hour task limit. For a like-for-like six-hour comparison, Carbon scored 53.5%, ahead of Argon at 39.5% and Claude Opus 5.5 at 34.5%. These evaluations were run without model guardrails, and a test score does not establish real-world safety.
Exploit validation benchmark using six-hour caps for all models shown. Source: Google DeepMind.
Finding bugs is only half the job
Carbon also reached 71.0% on CWE-bench, a benchmark for fixing code weaknesses, versus 68.0% for Argon, GPT-6 Astra and Grok 4.7. The published comparisons use different agent tools, so the results reflect the combined testing setup, not just each model in isolation.
Code patching benchmark, including the final supplied comparison. Source: Google DeepMind.
Google says introductory API pricing will be $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period. It has not given a general-release date. For now, the bigger test is whether Carbon can help defenders find, validate and fix real vulnerabilities before similarly capable tools become widely available.
Get stories like this in your inbox every week. Subscribe to The Week in Perspective.