Anthropic Says Claude Wrote Over 80% of Its Merged Code—Review Is the Bottleneck

In an Anthropic Institute analysis of internal engineering data, the company says that as of May 2026 Claude authored more than 80% of the code merged into its codebase; it also says the typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024, while warning that code volume overstates the true productivity gain and identifying human review as a new bottleneck.

This is a more precisely scoped update to the looser claims that circulated during 2025. The central point remains valid: Claude now performs most implementation work entering Anthropic’s repositories, but people still select the goals, evaluate the results and decide what reaches production. For creators who build software, that distinction matters more than the headline percentage.

What the new figure actually measures

The metric covers lines of code attributed to Claude and merged into Anthropic’s production codebase. It does not describe Claude’s share of every historical file, all engineering activity or the total intellectual work behind a change. Requirements, architectural choices, testing decisions and deployment responsibility are not captured by a line count.

The scope is also narrower than informal estimates that may include scripts, experimental work and other code outside production. Anthropic acknowledges gaps in its attribution system, while noting that some unattributed lines are generated artifacts rather than code typed by people. The result is therefore a company measurement with stated limitations, not an independent audit or a universal benchmark.

“Merged” is an important qualifier. Generated code has passed far enough through Anthropic’s workflow to enter a repository, but that does not mean Claude independently conceived, validated and approved the work. The metric establishes authorship of implementation, not ownership of the full development process.

How the claim changed over time

The earlier public statements measured different parts of the workflow. In WIRED’s May 23, 2025 account of Anthropic’s developer conference, chief product officer Mike Krieger said Claude Code was writing more than 70% of the company’s pull requests. A pull request is a proposed set of changes, so its share cannot be treated as the same thing as the proportion of merged lines.

The subject later expanded from routine company development to work on Claude itself. In Axios’s September 17, 2025 account of the AI+ DC Summit, CEO and co-founder Dario Amodei described Claude as writing the “vast majority” of code for future versions of the model; Amodei and co-founder Jack Clark also acknowledged that it could not manage the entire development process.

These statements are related but not interchangeable. One concerns pull requests, another concerns code for future Claude systems, and the newer internal measurement concerns attributed lines merged into the wider Anthropic codebase. Treating them as a single continuous percentage would conceal changes in both methodology and subject.

Code authorship is not autonomous engineering

Producing an implementation is only one layer of software development. A team must decide which problem deserves attention, translate an ambiguous need into a testable objective, account for dependencies and security constraints, and determine whether the result is safe to deploy. A model may write most of the resulting lines without controlling any of those decisions.

Anthropic’s data draws the clearest boundary around goal selection. Claude can increasingly receive a desired outcome and determine a method, but larger gaps remain when the task requires deciding what should be built in the first place. That is why a high authorship rate can coexist with continued reliance on experienced engineers.

The working relationship is better described as supervised delegation. An agent can inspect a repository, edit several files, execute tests and revise its solution, while an engineer supplies context and evaluates whether the apparent success is genuine. Technical knowledge remains necessary because a change can pass narrow tests while violating an architectural assumption or creating a problem elsewhere.

Faster generation moves the bottleneck to review

When implementation becomes faster, the scarce resource shifts toward evaluation. Engineers must inspect a larger volume of plausible changes, preserve consistency across the codebase and decide how much testing is appropriate for each risk. If review capacity does not grow with generation capacity, the queue simply moves downstream.

This also explains why the increase in merged lines cannot be read as an equivalent increase in business value. Code volume can represent useful features and repairs, but it can also include generated tests, verbose implementations and additional maintenance obligations. Customer outcomes, reliability and delivery time are different measurements.

Anthropic is an unusually favorable environment for its own coding system. Its engineers work close to the models, can use internal capabilities and have strong incentives to redesign workflows around agents. The company’s experience demonstrates what intensive adoption can produce inside a frontier AI laboratory; it does not supply a productivity multiplier that another studio, startup or independent developer can assume.

What the result means for software creators

The consequential change is not merely that developers type fewer lines. Agentic coding allows implementation to proceed in larger delegated units: the system can gather repository context, use development tools and iterate on feedback instead of suggesting an isolated snippet. That expands what one person can attempt, while increasing the amount of generated work that person must judge.

Human contribution consequently becomes less visible in the final code. Problem framing may be embedded in a prompt, architectural judgment in constraints, and expertise in the decision to reject a superficially convincing result. A code-authorship statistic does not capture those inputs, even though they determine whether the output is useful.

The updated figure therefore supports a narrower and more defensible conclusion than the claim that programmers have been replaced. Claude performs most of the measured implementation entering Anthropic’s codebase, while human judgment remains upstream in goal selection and downstream in review. The company has changed who writes the lines; it has not removed responsibility for deciding what those lines should do.

Also read:

• Anthropic’s “When AI Builds Itself” Report: 80% of Their Code Is Now Written by Claude
• O2 Says Its Daisy AI Averted Over 1,000 Scam Calls
• Anthropic Just Shipped Its Own “OpenClaw” Faster Than OpenAI — Meet Dispatch for Claude Cowork