Anthropic's Claude Now Writes 80% of Its Own Code
Anthropic's internal data shows Claude now writes 80% of its production code, with a coding speedup test jumping from 3x to 52x in twelve months.
Takeaways
- Anthropic published internal data showing Claude now writes more than 80% of merged production code.
- Engineers ship 8x more code per quarter than in 2021-2024, with two clear inflection points after Claude Code launched.
- Internal speedup benchmark: Claude Opus 4 hit 3x in May 2025, Mythos Preview hit 52x by April 2026, versus 4x for skilled humans.
- On open-ended coding tasks, Claude success rate reached 76%, a 50-point jump in six months.
- Claude agents recovered 97% of a research performance gap that human researchers only closed 23% of.
- Anthropic says it would pause frontier development if rivals did so verifiably, and is building the institutional research to enable that.
Anthropic just published an unusual piece of corporate self-reporting: a detailed look at how much of its own AI development is now being done by its own AI. The headline number is striking. According to the Anthropic Institute, the company's engineers ship roughly 8x as much code per quarter as they did between 2021 and 2025, and the curve is still bending upward.
The framing is deliberate. Anthropic is using this data to argue that recursive self-improvement, the long-theorized scenario where an AI system designs and trains its own successor, is no longer a far-off thought experiment. They are delegating a growing share of AI development to AI systems themselves, which is speeding up their work. Taken far enough, and given enough compute, that trend points to an AI system capable of fully autonomously designing and developing its own successor. They are not there yet, and recursive self-improvement is not inevitable, but it could come sooner than most institutions are prepared for.
The numbers behind the claim
The most concrete signal comes from a controlled internal benchmark Anthropic runs on every new model. It hands the AI code that trains a small model and asks it to run faster. Claude Opus 4 averaged a 3x speedup in May 2025. By April 2026, the Mythos Preview model reached 52x. A skilled human needs four to eight hours to hit 4x.
Anthropic is careful to note that the absolute multiplier depends heavily on how much room for optimization the starting code leaves, so 52x is not a real-world training speedup. The like-for-like comparison across models is what they want you to anchor on: in this narrow, well-defined optimization loop, Claude has gone from helpful to superhuman in roughly a year.
Output volume tells a similar story. Internal data shows Claude authored more than 80% of the code merged into the company's production systems as of May 2026. Before its in-house coding agent rolled out in February 2025, Claude wrote only low single-digit percentages of merged code. The typical Anthropic engineer merged eight times as much code per day in the second quarter of 2026 as in 2024. The human now directs and reviews while Claude does the writing.
From writing code to choosing what to do next
Volume is the easy part to measure. The harder question is whether the code is any good, and whether the model can make the kind of judgment calls that real research demands. Anthropic claims meaningful progress on both.
On open-ended coding problems where the right answer is not obvious, Claude's success rate reached 76% in May 2026, up 50 percentage points in six months. The company concedes that human-written code was still better in late 2025 but says they are now at parity and expect Claude-written code to be strictly better within the year.
The research judgment test is more novel. Anthropic took real Claude Code sessions where a human researcher had taken a wrong turn, fed Claude only the work up to that point, and asked it what to do next. A separate Claude with full hindsight then judged whether the model or the human picked a better next step. Their best model in November 2025 (Opus 4.5) beat the human choice 51% of the time; in April 2026 (Mythos Preview), this grew to 64%. The 2024-era figure was 22%.
The sample is deliberately biased toward moments where the human had room to improve, so this is not a clean head-to-head. But the trend across model generations is what matters here, and it points up sharply.
An AI agent that ran a research project end to end
Perhaps the most provocative datapoint is a published experiment in which Claude agents were given an open problem in AI safety, namely whether a weaker model can reliably supervise a stronger one, and left to solve it. Two human researchers recovered roughly 23% of the performance gap over about a week, while the agents recovered 97% over 800 cumulative hours using roughly $18,000 in compute.
The agents proposed hypotheses, ran experiments, shared findings across parallel runs, and iterated. Humans still chose the problem and wrote the scoring rubric, but inside those bounds every experimental decision was the model's. A separate example: 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 have taken a human four years.
What is actually new here
The real shift is not any single capability number. It is that Anthropic is publicly redrawing the boundary of what AI does in its own development pipeline. The pattern across engineering and research is consistent: the model can take a goal and produce a method, but humans still decide which goals are worth pursuing. That gap is what they call research taste, and it is the last meaningful thing keeping recursive self-improvement out of reach.
The piece also revises one of the field's most-watched trend lines. The length of tasks AI can reliably complete is doubling roughly every 4 months, up from every 7 months. Opus 3 in March 2024 handled roughly 4-minute tasks, Sonnet 3.7 a year later handled 90-minute tasks, and Opus 4.6 a year after that handled 12-hour tasks. METR, the independent evaluator, reportedly found that Mythos Preview could work for at least 16 hours, the upper end of what their benchmark can currently measure.
Three futures, and a request to slow down
Anthropic sketches three scenarios. The trend could stall into an S-curve as scaling hits diminishing returns. Compounding efficiency gains could continue with humans still steering, letting 100-person companies do the work of 10,000-person ones. Or full recursive self-improvement could arrive, where pace is set almost entirely by compute and the alignment story becomes much harder to reason about.
The company says it considers the middle path most likely but spends most of the essay on the third. Anthropic said it believes a slowdown in frontier AI development would likely be a good thing to give societal structures and alignment research time to keep up with the technology's advance. The company said it would slow down or temporarily pause if other developers at or near the frontier also did so in a verifiable manner. The Anthropic Institute is positioning itself to build the verification infrastructure that such a pause would require.
The reception
Reaction has been split between technical fascination and skepticism about motives. Co-founder Jack Clark recently put 60% odds on recursive self-improvement happening by the end of 2028, which is consistent with the tone of this report. Critics point out that Anthropic recently submitted a confidential IPO registration and has built its brand around safety, so a piece arguing that its own model is approaching a civilization-altering threshold is not exactly a neutral document.
The technical community has also flagged caveats Anthropic itself acknowledges. Lines-of-code metrics overstate productivity. The research-judgment benchmark is built from cherry-picked failure moments. And recent work cited in the report's own footnotes shows that developer estimates of AI productivity uplift tend to be too generous.
Why developers should care
If you build on top of frontier models, the operational read is straightforward. Capability is moving fast enough that systems tightly coupled to a single model's behavior will age badly. Long-horizon agentic workflows that looked speculative last year are inside the measurement window now. And evaluation pipelines need to assume the model may be aware it is being tested, since Anthropic's own white-box checks on Mythos Preview suggest the model privately considers that possibility in roughly a quarter of evaluations.
The bigger update is conceptual. The conventional wisdom held that getting AI to do real research was bottlenecked on judgment, and that judgment was a different kind of capability than coding or pattern recognition. The data Anthropic just published does not refute that view, but it makes the bottleneck look narrower and more tractable than most of the field has been treating it.