Gemini 4 Carbon tops Artificial Analysis Index at 59 but launches gated to cyber defenders

Sundar Pichai announced Carbon on LinkedIn (42K+ likes): 'I did say we'd be releasing our models periodically.' He highlighted Carbon's #1 ranking on the Artificial Analysis Intelligence Index v5 and its lead on cybersecurity evaluations:

- 88.2% on SEC-bench Pro (vulnerability finding).

- 54.8% on ExploitGym (exploitation).

- 71.0% on CWE-bench (patching).

Google also cited significant long-horizon task gains over Argon. Genie 4, released alongside, targets persistent interactive world generation.

The distribution model is the story. Because Carbon is strong at finding and exploiting vulnerabilities, Google restricted initial access to the Fairwind Program for vetted cyber defenders and gave no date for general developer access. Perplexity CEO Aravind Srinivas shared early evals calling it the 'strongest model that we've seen so far for long-form research', with a new state of the art on the WANDR benchmark. He said it will come to Perplexity Pro and Max users once they have enough tokens.

Gating split the community. A Hacker News thread (850+ points) called it paternalistic and questioned benchmark claims nobody can independently test. Mistral took the opposite route this week, publishing open weights for its cyber-capable Large 4. A Mistral researcher argued that some of the most serious recent attacks came from closed models anyway.

Watch for:

- The timeline for general API availability.

- Independent evaluations once access widens.

- Whether gated cyber-first release becomes the norm for frontier labs, as OpenAI's Daybreak Red on GovCloud suggests.