Anthropic is rolling out a new way to answer the question every engineering leader has been quietly asking about AI pair programmers: “Is this thing actually helping us ship more software, or just making dashboards look pretty?”

The company has launched contribution metrics for Claude Code, a new set of analytics tied directly to GitHub that tracks how much of your team’s real, merged work is done with AI assistance. Instead of vague “time saved” estimates or lines-of-code vanity stats, Anthropic is leaning on hard signals: pull requests shipped and code actually committed with Claude’s help.

Inside Anthropic, where Claude Code has been dogfooding for a while, those numbers are already eye‑catching. The company says that as internal adoption grew, it saw a 67% increase in pull requests merged per engineer per day, and estimates that 70–90% of code across teams is now written with Claude Code assistance. That internal story lines up with broader survey data Anthropic has shared about Claude usage, boosting average productivity by around 50%, with a subset of power users more than doubling their output.

The new metrics product is essentially Anthropic’s attempt to package that internal experience into something an engineering org can monitor without building a bespoke data pipeline. Once you connect Claude Code to your GitHub organization via the Claude GitHub App, the system starts matching Claude Code sessions to specific commits and pull requests. Only the changes where Anthropic has high confidence Claude was involved are labeled as “assisted,” which is meant to avoid over‑crediting the AI every time a developer happens to have the tool open.

In practice, contribution metrics surfaces three main views: how many pull requests were merged with versus without Claude Code, how many lines of code were committed with and without AI assistance, and per‑user contribution data that shows adoption patterns across your team. All of this lives in the existing Claude Code analytics dashboard that Team and Enterprise customers already get, so there’s no extra SaaS, just a GitHub integration and some admin toggles in the Claude settings.

For leaders, the pitch is that this plugs into the metrics they already care about. Anthropic explicitly calls out that contribution metrics are meant to sit alongside things like DORA metrics, sprint velocity, or other internal KPIs, not replace them. You can imagine a head of engineering looking at PR throughput, lead time, and incident rates, and then layering on “what changed after we rolled out Claude Code to this squad?” rather than debating survey responses about whether developers feel “20% faster” this quarter.

It also reflects where the whole AI‑coding‑tools market is heading. Other players and platform vendors are starting to talk about PR cycle times, PRs per developer, and lead‑time deltas between AI and non‑AI teams as more credible indicators of impact. Anthropic’s move is to wire that kind of view directly into the tool, instead of forcing teams to bolt it together themselves or rely on one‑off ROI analyses.

From a rollout standpoint, the feature is in public beta and limited to Claude Team and Enterprise plans for now. Enabling it is fairly straightforward: a GitHub admin installs the Claude GitHub App for the organization, a Claude workspace owner switches on GitHub Analytics in the Claude Code admin settings, and then someone authenticates against the GitHub org. After that, the metrics start populating as developers use Claude Code in their normal workflows, with Anthropic’s docs filling in details on setup and how to interpret the graphs once they’re live.

There are still the usual caveats: pull requests are an imperfect proxy for “velocity,” and more PRs or more lines of code do not automatically equal better software. But Anthropic is leaning into the idea that if AI assistants are going to become part of the engineering stack, they need to be held to the same standard as any other tool: show up in the numbers that matter, or get out of the way. Contribution metrics is Anthropic’s answer to that challenge, and, for teams already experimenting with AI pair programming, it gives them something more concrete to point to than yet another anecdote about “it feels faster.”