Gemini 4 Argon Is Powerful and Deliberately Hard to Reach
Google gave its best new model to cyber defenders first. That tells you who they're afraid of.
Gemini 4 Argon shipped October 1 with a million-token output and a reported 68% score on a vulnerability benchmark. The interesting part isn't the size, it's the gate: early access went to vetted cyber defenders through the Fairwind program, not the public API.
That's a recognition that a model good at finding and patching flaws is also a model good at finding them elsewhere. Google is rationing capability by trust, which is rare and probably wise.
For the rest of us, Argon reaches the Gemini app on October 14 for paid tiers. I'd use it for the boring, enormous review jobs -- a whole codebase or contract -- not for clever prompts.
The lesson for builders: the most capable models are starting life inside permission boundaries. Expect that to become the norm, not the exception.