Anthropic: Claude Wrote 80% of Merged Production Code

Anthropic reported that its Claude model authored more than 80% of the code merged into the company’s production systems by May 2026, following an in-house coding-agent rollout in February 2025. The company said typical engineers merged about eight times as much code per day in the second quarter of 2026 as they did in 2024.

The Anthropic Institute published a report that combined internal merge data with public benchmarks. Before the coding agent was deployed, Claude accounted for low single-digit shares of merged code; by May 2026 that share exceeded 80%. Human engineers shifted toward directing work and reviewing model-produced code.

Anthropic ran repeatable tests that measure model performance on training and research tasks. In a benchmark that asks a model to speed the training of a small model, Claude Opus 4 averaged roughly a 3x speedup in May 2025. By April 2026, the Mythos Preview model reached a 52x speedup on the same test. The report states an experienced human typically needs four to eight hours to achieve about a 4x improvement on that task.

On a research-judgment test that asks a model to select the better next step in a development session, Mythos Preview chose the better option 64% of the time, compared with 51% for Opus 4.5 in November 2025.

An excerpt in the report read: ‘Claude-written code was somewhat worse than human-written code at Anthropic in late 2025, is roughly at parity today, and we expect it to be strictly better within the year.’ The report also noted Claude has not demonstrated the ability to choose which research problems matter most.

Anthropic has filed a confidential IPO registration and has emphasized safety as part of its public messaging. The company disclosed the data through the Anthropic Institute and social posts on June 4, 2026, highlighting changes to its internal workflow and the benchmark results.