Parity, According to Whom?

In June 2026 Anthropic's institute published "When AI Builds Itself," and buried in the recursive-self-improvement victory lap is a quiet grading curve. Claude-written code was "somewhat worse than human-written code" in late 2025, the report says. It "is roughly at parity today." And Anthropic expects it to be "strictly better than human within the year."¹ That is a remarkable thing for a company to say about its own product. It is also, depending on which scoreboard you read, either true or beside the point.

Start with what the report actually measures. More than 80% of the code Anthropic now merges into its own codebase is authored by Claude, up from low single digits before Claude Code launched in early 2025. In the second quarter of 2026, the typical Anthropic engineer was merging roughly eight times as much code per day as in 2024.¹ Anthropic attaches a caveat to that figure, and it belongs here rather than buried. Lines of code, the company writes, "is an imperfect measure, as it measures quantity over quality," and the eightfold figure "is almost certainly an overstatement of the true productivity gain."¹ Those are real numbers, and they describe one specific thing very precisely: how much Claude-generated code makes it into Anthropic's repositories, judged by Anthropic's own engineers, on Anthropic's own work.

"Parity" is doing a lot of quiet work in that sentence. Parity at what, measured by whom, on whose codebase. Anthropic's answer is "the code we choose to keep." The independent receipts answer a different question, and they have been answering it for a year.

The Other Scoreboard

LinearB analyzed 8.1 million pull requests across more than 4,800 organizations and found that AI-generated PRs are accepted at 32.7%, against 84.4% for human-written code.² Read those two rates next to Anthropic's claim. Inside Anthropic, 80% of merged code is Claude's. Across the wider industry, two-thirds of what AI proposes gets rejected before it merges. Both can be true at once, because they are not measuring the same act. One counts code that survived. The other counts code that was offered.

LinearB adds a wrinkle worth keeping: AI PRs wait 4.6 times longer to be picked up by a reviewer, then move twice as fast once someone finally starts.² The bottleneck is not the typing. It is the queue of humans deciding whether the typing was any good.

Then there is reliability, which is the part "parity" tends to skip. DORA's 2024 research found that even as 75.9% of teams adopted AI, software delivery throughput slipped about 1.5% and delivery stability fell 7.2%.³ DORA's 2025 edition is more generous to the tools, revising the throughput relationship upward, but stability stayed stubbornly negative: more AI adoption still tracks with less stable software in production.⁴ Parity in a code review is not the same as parity at 3 AM when the system pages.

One scoreboard counts code that survived review; the other counts code that shipped and held in production.

The Felt Sense and the Stopwatch

The widest gap is between how fast AI coding feels and how fast it is. METR ran a randomized controlled trial with 16 experienced open-source developers in 2025 and clocked them 19% slower on real tasks when allowed to use AI tools. The same developers believed AI had sped them up by about 20%.⁵ A 39-point swing between the stopwatch and the gut.

In fairness to the tools, METR walked its own headline back. A February 2026 update, after recruiting a broader pool, estimated a modest speedup for the original developers and a slight slowdown for new ones, while calling the whole thing "only very weak evidence" and announcing a methodology redo.⁶ That reversal does not rescue the parity claim so much as underline the problem with it. The single most-cited measurement of AI coding productivity could not hold its own sign stable across seven months. When the rigorous independent study is that uncertain about direction, a vendor's confidence that its code is "strictly better within the year" is a forecast, not a finding.

Same Word, Two Games

None of this makes Anthropic's number false. Eighty percent of merged code being Claude's is an extraordinary internal milestone, and the company deserves credit for publishing the grading curve instead of just the trophy. The same report that brags about recursive self-improvement also calls for a verifiable mechanism to slow or pause frontier AI development across labs.¹ The brag and the brake, in one document. That is more honesty than most of the field offers.

The trouble is the word. "Parity" implies a shared yardstick, and there isn't one. Anthropic measures what gets merged into its own codebase, while the receipts measure acceptance rates, production stability, and stopwatch time. A model can be at parity on the first scoreboard while the second one says AI code is rejected twice as often, ships less stable software, and slows experienced developers down. Whose parity wins is not a technical question. It is a question of which number you were already paid to believe.

· · ·

A year is a short forecast and an easy one to audit. The receipts are public, the methodologies are published, and the deadline is on the record. We will be reading the box score, not the press release.