Anthropic: Claude Wrote 80% of Code Merged into Production

Anthropic Institute released internal data showing Claude authored more than 80% of the code merged into Anthropic’s production systems by May 2026. The data shows engineers merged eight times more code per day in the second quarter of 2026 than in 2024.

The increase followed deployment of an in-house coding agent in February 2025. Before that rollout Claude accounted for low single-digit shares of merged code; by May 2026 its contribution exceeded 80%. Engineers now focus on directing and reviewing work while Claude generates the bulk of code that reaches production.

The institute ran controlled tests to measure how models speed up training tasks. In a standard test that provided code to train a small model and asked the model to make training faster, Opus 4 averaged about a threefold speedup in May 2025. By April 2026 a Mythos Preview build produced a 52-fold speedup. The company noted a skilled human typically needs four to eight hours to reach a fourfold improvement on the same task.

The report measured research judgment by presenting models with a session in which a researcher took a suboptimal turn and asking what the next step should be. Mythos Preview selected the better next step 64% of the time, up from 51% for Opus 4.5 in November 2025. The report states: “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.”

Anthropic characterizes the results as an early signal of recursive self-improvement but cautions Claude has not demonstrated the ability to choose which research problems matter most. The company emphasizes continued human oversight and review of model outputs.

Anthropic filed a confidential IPO registration and promotes safety research as part of its brand. The report notes faster internal development could affect timelines for new product features and the company’s competitive position. It also says the trend has drawn interest in autonomous AI agents across other sectors, including crypto, where agents are used to execute trades and perform on-chain tasks.

The company provided the data to show how automated coding and research assistance can change engineering workflows and to raise governance and risk questions as models assume more development work. The report adds that whether gains continue to accelerate or flatten will determine how soon AI systems begin reliably building more capable successors.