Unreleased models don't count
Many see announcing Carbon without access as earnings-driven marketing, and say tiny Pro quotas on Argon make Google's models effectively unusable anyway.
Gemini hallucinates badly
Numerous users report Argon fabricating products, transcripts, and facts, and calling it the worst frontier model for hallucination and weak tool use despite strong benchmarks.
Artificial Analysis index looks suspect
Commenters note index revisions keep coinciding with big-lab releases and dropping open-weight models, though others argue new unsaturated, private benchmarks simply expose benchmaxxing.