When AI Builds Itself: Reading Anthropic's Recursive Self-Improvement Essay
As of May 2026, more than 80% of the code Anthropic merges into its own codebase is written by Claude, not by humans. A year earlier it was in the single digits. That one number is the whole story of Anthropic's essay When AI Builds Itself — AI is crossing from a tool that helps people write code into a system that increasingly writes itself.
I came to the essay through a video reaction to it, then went and read the primary source — and I'm glad I did, because the video got a few important things wrong. So this post does two things: it lays out what Anthropic actually claims (with the real figures), and it flags where the popular "doom" framing diverges from the text.
Sourcing note: every statistic below is quoted from Anthropic's own essay, linked above. Where I mention "the video's take," that's a third-party reaction — clearly a different source with a different agenda.
The Five Stages: Humans Getting Abstracted Away
Anthropic frames the last five years as a steady removal of humans from the act of coding:
- Building the first Claude (2021–2023) — engineers typing code by hand.
- Chatbots (2023–2025) — people paste snippets out of a chat window into their editor.
- Coding agents (2025–2026) — you describe a task; an agent writes whole files.
- Autonomous agents (today) — one prompt fans out to sub-agents that run code and train models in parallel.
- Closing the loop — the future stage: AI designs, builds, and trains its successor with no human in the loop.
Worth noting: stage four is labeled "today," not "near future." Anthropic is saying the autonomous-agent era is already here, and only the final recursive step remains.
The Acceleration Is Real — and Measured
The compounding is the part that's hard to argue with, because it's benchmarked. The length of task an AI can reliably complete on its own is doubling roughly every four months — up from every seven. In concrete terms, straight from the essay:
| When | Model | Task length it could handle |
|---|---|---|
| March 2024 | Claude Opus 3 | ~4 minutes |
| March 2025 | Claude Sonnet 3.7 | ~1.5 hours |
| April 2026 | Claude Opus 4.6 | ~12 hours |
| 2026 (projection) | — | weeks-long tasks "could come into range this year" |
On CORE-Bench — which tests whether an AI can read a research paper, write the code, and reproduce the results — systems went from ~20% success in 2024 to saturating the benchmark fifteen months later.
Inside Anthropic Right Now
Beyond the headline 80%, two internal numbers stand out — and this is exactly where the video's framing slipped:
- 8× more code per day. In Q2 2026, the typical engineer merged 8× as much code per day as in 2024. Anthropic itself adds the caveat that lines of code measure quantity over quality, so this is "almost certainly an overstatement."
- ~4× more output. Separately, the median engineer estimated producing ~4× as much output using an internal model preview called Mythos.
The video collapsed these into "8× code but only 4× productivity, therefore AI code is half as valuable." That's a tidy story, but it isn't what the essay says — the 8× and 4× measure different things (code volume vs. self-estimated output), and Anthropic flags the 8× as inflated. It's a real signal of acceleration, not an admission that half the code is junk.
The sharpest data point: on a code-optimization task, Claude Mythos Preview hit a ~52× speedup over the starting code (April 2026), versus ~3× for Claude Opus 4 a year earlier — where a skilled human researcher needs four to eight hours just to reach 4×.
The Real Bottlenecks: Judgment, Then Compute
If AI writes most of the code, why are humans still here? Because execution and direction are different skills. AI is strong at execution and still weak at deciding what's worth building — coming up with novel research directions, choosing which experiments to run, judging results. Anthropic notes that human code review has itself become a new bottleneck.
The deeper constraint is physical. The essay argues the binding limit on AI progress may end up "in the supply chain, not the model" — that pushing and spreading the frontier could need more energy and compute than currently exists. Whoever can afford the most compute and electricity sets the pace.
Here's a correction worth making: the video dramatized this as a coming "permanent underclass." That phrase appears nowhere in Anthropic's essay. The compute-and-capital constraint is real and worth taking seriously, but the dystopian label is the commentator's, not the company's.
The "Pause" Question — and Where the Critique Misfires
This is the part most worth getting right. Anthropic writes that if it were genuinely possible to slow this technology down to buy society more time, that "would likely be a good thing." A common critique — the one in the video — calls this self-serving fear-based marketing: the leader in a race benefits from telling everyone else to slow down.
It's a fair instinct to be skeptical of any frontier lab's safety messaging. But the critique misreads the actual commitment. Anthropic does not propose a unilateral pause. The essay says it would slow or pause only if other frontier developers also did so in a verifiable manner, and that a real slowdown "would require multiple well-resourced labs at or near the frontier, in multiple countries, agreeing to stop under the same conditions." That's close to the opposite of "I'll stop and hope competitors do too" — it's "none of this works unless it's collective and verifiable."
My own read: you can hold both thoughts at once. A market leader's safety advocacy deserves scrutiny, and a conditional, multilateral, verification-based pause is a far more defensible position than the strawman version going around. Criticize the essay for what it says, not for the more cynical thing it's easier to dunk on.
Frequently Asked Questions
What is recursive self-improvement in AI?
Recursive self-improvement is the point at which an AI system can autonomously design, build, and train a more capable successor with little or no human involvement — each generation improving the next. In Anthropic's framing it's the final "closing the loop" stage, where the main limit on progress becomes raw compute rather than human engineering effort.
How much of Anthropic's code is written by AI?
According to Anthropic's essay, more than 80% of the code merged into its codebase as of May 2026 was authored by Claude — up from single digits a year earlier.
What is CORE-Bench?
CORE-Bench tests whether an AI can read an existing research paper, write the necessary code, and successfully reproduce the paper's results. AI went from roughly 20% success in 2024 to saturating the benchmark about fifteen months later.
Did Anthropic call for a unilateral pause on AI?
No. Anthropic said it would slow down or temporarily pause only if other frontier labs did the same in a verifiable way, and that a meaningful pause would require multiple well-resourced labs across multiple countries agreeing under the same conditions.
Where I Land
The numbers in this essay are not hype — they're measured, caveated, and pointed in one direction. AI is moving from writing code to directing the work, and humans are sliding up the stack from coder to reviewer to director. The honest uncertainty is whether AI ever masters taste — novel research and judgment — because that's the only thing keeping people in the loop.
What I'd push back on isn't Anthropic's data; it's the secondhand framing that turns a careful, hedged essay into a tidy doom narrative. Read the primary source. Then decide what to be worried about — there's plenty in the real version without inventing an underclass.