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, but developer complaints began within a day. Developer Theo Browne said on October 14 that Argon was unusable in his own repositories because it confidently invented explanations of how existing systems worked, and that more than half of its factual claims were wrong in almost every test he ran. That result was one developer's report, not an independently replicated error rate.

Other users complained about speed and about Pro allowances being exhausted quickly, while Google pointed to a supervised 11-hour run on a roughly 205,000-line Python repository with 16,892 tests. Reviewers accepted two of six proposed optimisations there, with reported gains in session preparation and cold imports. The brief argues the missing comparison is net engineering work: how often code is described correctly, how many changes survive tests, and how much human checking and rework a task takes.