Anthropic

Claude writes roughly 80% of the code merged into Anthropic's monorepo, where daily changes rose from about 2,500 in May 2026 to more than 30,000 in September. Sachin Malhotra traced what that did downstream: ten times as many tests at near-constant headcount, twenty-five times as many CI jobs in six months, and a test impact analysis service that had to be rebuilt around an in-memory store. One engineer shipped that rebuild in three weeks, against roughly a quarter a year earlier.

The company treats itself as an early indicator. Four economists from its Economics team reported on 16 October 2026 that junior technical roles fell from 21% to 14% of technical departments since early 2025, and from 27% to 9% of hires. Agent hours passed human hours in March 2026. Non-technical employees went from two code-agent sessions a month to ninety. Human attention became the constraint: about 80% of agent session time is spent waiting for a person, and "meta-work" grew from 10% to 22% of sessions.

The same reflex shows in review. The Claude Code team swapped "did a human read this?" for "what control should catch what can go wrong," and ruled that an agent's approval never substitutes for a required human one.

Safety runs on the same logic of arbitration. Claude Fable 5.1 and Claude Mythos 5.1 are one underlying model at two guardrail levels, with Mythos reserved for trusted-access programs. One recurring charge against the company is that this amounts to a permission regime selling cognition as infrastructure.