Anthropic engineers ship 8X more code than last year, with 80% authored by Claude

Anthropic’s engineers are merging roughly eight times as much code per quarter as they did in 2024. The kicker: over 80% of that code was written not by the humans, but by Claude.

The figures, disclosed publicly in June 2026, paint a picture of a company that has effectively turned its own product into its most prolific developer. Before the launch of Claude Code in February 2025, AI-generated code contributions at Anthropic sat in the low single digits. Eighteen months later, Claude is writing four out of every five lines that make it into production.

The numbers, and why Anthropic doesn’t fully trust them

Anthropic itself described the 8x productivity claim as “almost certainly overstated,” labeling lines-of-code as a vanity metric.

Internal metrics showed that Claude’s success rate on open-ended coding tasks climbed from approximately 26% to 76%.

One engineer reportedly went five months without writing a single line of code personally, relying entirely on Claude for all code output.

What changed since February 2025

The inflection point was the launch of Claude Code in February 2025. Before that tool existed, AI contributions to Anthropic’s codebase were negligible. The tool essentially gave engineers a way to delegate entire coding tasks to Claude rather than just using it for suggestions or snippets.

The Q2 2026 data reflects the cumulative effect of that transition. The daily volume of merged code roughly multiplied by eight compared to 2024 benchmarks, suggesting that the bottleneck in software production shifted from writing code to reviewing it.

The broader industry implications

The “almost certainly overstated” caveat matters more than most people will acknowledge. Shipping more code doesn’t automatically mean shipping better software. It doesn’t mean fewer bugs, more reliable systems, or happier users.

The fact that Anthropic, the company most invested in this narrative succeeding, is the one cautioning against taking the numbers at face value should tell you something. The productivity gains are real, but measuring them honestly remains an unsolved problem.

The 26% to 76% success rate on open-ended tasks suggests Claude isn’t just getting faster at doing simple things. It’s getting meaningfully better at doing hard things.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.