Anthropic defaults Claude Code to Auto Mode, boosting pull requests by 25%

Starting August 14, Anthropic will ship Claude Code with Auto Mode switched on by default for Pro, Max, and Team subscribers. Enterprise customers still need to opt in manually. The change means the AI coding assistant will execute most actions on its own, only pausing to ask for a human sign-off when a classifier flags something dangerous or irreversible.

The numbers behind the decision are hard to ignore. In tests with 1,053 paid users and internal red teams, Auto Mode matched or exceeded the safety record of manual approvals. Teams using it generated roughly 25 percent more pull requests — a productivity bump that makes the business case for itself. Anthropic also points to security gains: an independent audit by Trajectory Labs ran 72 attack scenarios ten times each. None of the 720 prompt injection attempts broke through Claude’s current models in Auto Mode. OpenAI’s GPT-5.6 Sol, tested in Codex Auto-Review mode, let 5.83 percent of attacks slip past.

Anthropic shared two internal war stories that likely accelerated the default rollout. In one incident, Auto Mode prevented Claude from posting confidential data to a public page. In another, it killed around 2,000 processes that were on track to disrupt ongoing GPU training jobs — exactly the kind of costly, hard-to-spot failure that keeps infrastructure engineers up at night.

The company isn’t charging for the tokens its safety classifier burns through, but the economics still tilt in its favor. More autonomous sessions mean longer sessions, and longer sessions mean more total token consumption. Revenue rises accordingly, whether or not that was the primary motive.

What worries me is the paradox Anthropic itself acknowledges. The blog post still recommends that developers review Claude’s actions themselves for high-stakes production changes. That’s sensible advice, but it runs into a wall when the codebase you’re reviewing was largely written by an agent you didn’t closely supervise. The less often a developer intervenes, the harder it becomes to build the deep, intuitive understanding that makes oversight meaningful. This is the same tension we saw when self-driving cars first required drivers to keep their hands on the wheel — the system works until the human’s attention atrophies, and then you’re left hoping the classifier doesn’t miss something you’re no longer equipped to catch.