GOOGLE'S SMARTER AI TAKES LONGER TO BE WRONG

Google released Gemini 4 Argon on October 13 as its most capable model for complex coding and agentic work. Within a day, developers working in real codebases began reporting the opposite experience: Theo Browne said that in almost every test he ran, more than half of the model's factual claims about his existing systems were wrong, and called it unusable for coding in his repositories.

A separate developer reported that four hours of Argon High coding introduced 115 regressions caught by deterministic tests, followed by another four hours repairing the damage, and that Argon took roughly four times as long as Opus 5.5 High on the compared work. Both sets of numbers are the reporters' own claims; the full tasks, logs and matched conditions are not public.

Google's own counter-evidence is an 11-hour unattended run on a roughly 205,000-line Python repository, where reviewers accepted two of six proposed optimizations. The brief's conclusion: the reports establish a real backlash and conflicting experiences, but no matched test yet measures the metric that matters — how much verified engineering work survives, per hour of human cleanup.