Google Gemini 4 Carbon: New Frontier Model Tops AI Index, Starts With Cyber Defenders
Google has unveiled Gemini 4 Carbon, a new frontier model that the company describes as its strongest to date, with state-of-the-art cybersecurity and long-horizon capabilities. The model follows Gemini 4 Argon and, like its predecessor, is not yet available to the general public.
Benchmarks reported at launch
According to reports on the announcement, Carbon scores 59 on version 5.0 of the Artificial Analysis Intelligence Index, putting it ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57) on that ranking. Google also claims gains in agentic coding, computer use, science and multimodal tasks.
Cybersecurity is where Google stresses the biggest lead. The company reports the following results:
| Benchmark | What it measures | Carbon score |
|---|---|---|
| SEC-bench Pro | Finding vulnerabilities | 88.2% |
| ExploitGym | Exploiting vulnerabilities | 54.8% |
| CWE-bench | Patching vulnerabilities | 71.0% |
These figures come from Google and the benchmark operators, and independent testing will take time as access widens.
Limited access for now
Carbon is first being released to a small group of cybersecurity defence companies for pre-release testing. Google plans to add more trusted partners through its Fairwind Program before a broader rollout to developers, businesses and consumers. No date for general availability has been given. Argon followed the same path and has since reached the Gemini app and AI Mode in Search.
The staged release reflects a concern Google has voiced before: the abilities that help defenders find and fix software flaws could also be misused for attacks, so safeguards are being strengthened before wider access.
Pricing
Reports say Google will launch Carbon at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%.
Because the model is limited to select partners, most developers cannot test these claims yet. The next weeks should show how Carbon performs outside Google's own evaluations.
How to read the rankings
Leaderboard results depend on which benchmark is used. While the Artificial Analysis Intelligence Index v5.0 puts Carbon first at 59, another evaluator, Vals AI, ranks it second among 30 models on its own index with 70.36%. Differences like this are normal, because each index weights coding, reasoning and agentic tasks differently, and the newest tests are updated to be harder than earlier ones.
For businesses, the practical question is less the rank than the cost and reliability of the model in real workflows. Google says Carbon cuts shortcut-taking and improves honesty in multi-step tasks, but those are internal claims that outside teams have not yet reproduced.
