More Code Does Not Automatically Mean Better AI

1 One number, several meanings

Anthropic reports that its typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024, and that more than 80 percent of merged code was authored by Claude as of May 2026. The company itself flags the limits of the metric: lines of code measures quantity over quality, so the figure is, in its words, "almost certainly an overstatement" of true productivity gain.

2 Code volume is not model capability

Code output describes activity inside one company's development pipeline. It says nothing directly about how capable the resulting models are. Model capability is measured by benchmarks, independent evaluations such as METR's long-task measurements, and downstream performance, none of which are captured in a merge count.

3 Useful output, on self-report

For usefulness, the cited evidence is a March 2026 internal poll of 130 Anthropic researchers, with a median estimate of roughly four times the output with Mythos Preview. Anthropic expects the true uplift was lower, and cites METR research showing developer estimates of AI productivity can be overestimated. The company still finds a significant fraction of technical staff accomplishing core work multiple times faster with AI assistance.

4 Reliability, by a different measure

Reliability is tracked not by volume but by interventions: the rate at which staff correct, redirect, or take over mid-task from Claude has been falling steadily for a year, per Anthropic, including on open-ended tasks. Success on the most open-ended tier reached 76 percent in May 2026, up 50 percentage points in six months. Volume and quality move on separate evidence bases.

5 When quantity does signal something

One volume-adjacent claim carries specific content: in April 2026, Claude shipped over 800 fixes that reduced a class of API errors by a factor of one thousand, work the overseeing engineer estimated would take a human four years. Anthropic also notes a cost, that this surge is straining shared infrastructure, with GitHub on pace for roughly 14 billion commits in 2026 by mid-year, up from about one billion in all of 2025.

6 The bottom line

Eight-fold code volume is evidence of acceleration, not of better AI. Useful output rests on self-estimates Anthropic itself discounts, reliability on a falling intervention rate, and capability on benchmarks and external evaluations. Each claim needs its own evidence, and conflating them is the error to watch.