Google’s strongest AI model goes to cyber defenders first

In short: on 28 Oct 2026 Google announced Gemini 4 Carbon, the successor to Gemini 4 Argon, released two weeks earlier. The company calls it its strongest model, especially in cybersecurity and long tasks. The catch: for now only a subset of cyber defenders have access, and developers, businesses and consumers are to get it ‘as soon as possible’.

What Google reports

• Artificial Analysis Intelligence Index v5.0: 59 points and first place.
• Finding vulnerabilities (SEC-bench Pro): 88.2%, writing working exploits (ExploitGym): 54.8%, patching bugs (CWE-bench): 71.0%.
• Prompt injection attack success on the Gray Swan IPI test: down from 0.7% to 0.0% compared with Argon.
• More than twice the success rate for secure code and about half the unsupported claims compared with Argon, according to Google’s internal tests.

What it costs and who gets access

• Launch price: $2 per million input tokens and $10 per million output tokens, with 95% off cached input.
• After the launch period: $4 and $20. The Register notes this matches Claude Opus 5.5.
• Access: selected cyber defenders first, then more trusted partners through the Fairwind programme. Google has given no date for everyone else.

How to read this

Almost every number comes from Google, and some come from internal tests nobody outside can repeat. The exception is the Artificial Analysis index, run by an independent company. The limited launch also shows that a model good at finding holes is just as useful to attackers, and Google prefers to hand it to defenders first. I compare the strongest models in the frontier models hub.

Sources

Gemini 4 Carbon: our next frontier modelGoogle · Nov 1, 2026 · Primary sourceGoogle's Gemini 4 Carbon looks like a diamond in the roughThe Register · Nov 1, 2026