Claude Code adds metrics to track AI's impact on dev velocity

Engineering leaders have been spending millions on AI coding tools, but when the CFO asks whether anyone is actually shipping faster, the answer is usually a shrug and a few anecdotes.

Anthropic wants to put real numbers behind that question. The company just launched contribution metrics in Claude Code, currently available in public beta for Claude Team and Enterprise customers.

Anthropic engineers are already leaning on it heavily

To show what the numbers can look like, Anthropic shared data from its own engineering organization.

Since rolling out Claude Code internally, the company says pull requests merged per engineer per day jumped 67%. Even more striking, between 70% and 90% of all code written across Anthropic teams is now produced with Claude Code assistance.

Anthropic admits that pull requests alone do not tell the whole productivity story. Still, the company argues they are one of the closest proxies for shipping features and squashing bugs faster.

Conservative matching straight from GitHub

The new metrics plug directly into a team's existing Claude Code analytics dashboard, accessible to workspace admins and owners.

Setup requires installing the official Claude GitHub App and authenticating the organization account. There are no external data pipelines to configure.

Once connected, the dashboard tracks pull requests merged with and without Claude assistance, total lines of committed code, and per-user adoption patterns.

Anthropic says it calculates these numbers conservatively. The system matches Claude Code session activity to GitHub commits and PRs, counting work as assisted only when there is high confidence that Claude was involved.

Why it matters

AI coding assistants are moving from individual developer toys into corporate procurement line items, and line items require proof. For engineering managers, these metrics provide clear ammunition to justify enterprise seat costs or identify which team members need onboarding help.

We think Anthropic was smart to integrate directly into GitHub commits rather than inventing another proprietary workflow. It meets developers where their work already lives.

The catch, as any experienced engineering lead knows, is Goodhart's law. Pull request counts and lines of code are notoriously easy to game. If engineering orgs lean too heavily on these numbers as performance metrics, developers might start breaking simple tasks into multiple AI-assisted PRs, creating an avalanche of code review overhead without delivering real customer value.