TL;DR
The Anthropic Institute has published a report arguing that AI systems are now helping build AI at measurable scale, citing public benchmark gains and internal Anthropic data. The evidence includes Claude-written code and an April 2026 safety-research test, but Anthropic says recursive self-improvement has not arrived and may not arrive.
The Anthropic Institute has released evidence that AI is already speeding parts of AI development, including internal data on Claude-written code and an April 2026 agent-run research test, sharpening debate over whether recursive self-improvement could arrive before institutions are ready.
The report, titled When AI builds itself, argues that AI is now doing more of the practical work behind AI development: writing code, running experiments and producing results. It does not claim that AI systems are already designing successor systems without human direction. Anthropic says the remaining gap is research judgment: deciding which problems matter, which results are reliable and when an approach should be abandoned.
Anthropic points to public benchmark trends as one source of evidence. The report cites METR data showing that the length of tasks AI can reliably complete on its own is now doubling about every four months, compared with about every seven months earlier. In the examples given, Claude Opus 3 handled tasks of about four minutes in March 2024, Claude Sonnet 3.7 reached about 1.5 hours around March 2025, Claude Opus 4.6 reached about 12 hours in March 2026 and Claude Mythos Preview reached at least 16 hours in 2026.
The report also cites internal Anthropic figures, including claims that Claude has helped raise code output per engineer by about 8x and that more than 80% of merged code is written by Claude. The strongest internal example described is an April 2026 weak-to-strong supervision project in which agents generated hypotheses, ran tests, shared findings and iterated on an AI-safety problem. Anthropic says the agents recovered 97% of the measured gap between a weak-supervision floor and a strong-model ceiling, compared with about 23% for humans working for a week. The test used about 800 cumulative agent hours and roughly $18,000 in compute.
Why It Matters
The findings matter because they suggest AI labs may be approaching a point where the pace of AI development is set less by human labor and more by compute, tooling and access to experiments. If AI systems can move from executing research tasks to choosing useful research directions, model improvement could speed up across engineering, safety research and product development.
For readers, the practical issue is preparedness. Companies, regulators and research institutions may need to judge whether their evaluation methods, oversight processes and safety testing can keep pace with systems that increasingly assist in building the next systems.
Beyond Vibe Coding: From Coder to AI-Era Developer
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background
Recursive self-improvement refers to a loop in which an AI system helps design or improve a more capable successor, which then repeats the process. The Anthropic report presents that as a possible future path, not a current confirmed state.
The report divides frontier AI development into engineering and research work. In engineering, it says Claude can often take an underspecified goal and find a method. In research, it says Claude can execute well-specified experiments at or above skilled human performance, but humans still choose the question and judge the result.
The benchmark context also matters. SWE-bench, which tests real software bug fixes, moved from low single-digit performance to saturation over about two years, according to the source material. CORE-Bench, which tests reproduction of research papers, moved from about 20% performance in 2024 to saturation 15 months later.
“We are not there yet”
— The Anthropic Institute authors
NVIDIA DGX Spark™ – Personal AI Desktop Supercomputer – Desktop GB10 Grace Blackwell Chip
Supercomputer performance directly to your desk in a compact, energy-efficient design, enabling enterprise-scale AI and high-performance computing right…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What Remains Unclear
Several points remain unclear. Anthropic says the April 2026 agent-research result did not transfer cleanly to production-scale models, so it is not proof that agents can already improve frontier systems in a direct loop. The internal code and research metrics are also not fully checkable through public benchmarks alone. It is not yet clear whether AI can develop the research taste needed to pick valuable problems, reject misleading results and set a productive research direction without human control.
1set XIAOML Kit, Machine Learning Development Platform, Wi-Fi and BLE, ESP32-S3 Sense
CURRICULUM: Complete machine learning kit designed to Machine Learning Systems curriculum for structured, academic-grade learning experience.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What’s Next
The next tests will be whether AI agents can repeat these results on harder, production-relevant research tasks and whether independent benchmarks can measure longer, more open-ended work. Anthropic and other labs are likely to face growing pressure to publish clearer evidence about how much of AI development is being delegated to AI systems and where human judgment still governs the process.
AstroAI Digital Multimeter, TRMS 4000 Counts Volt Meter (Manual and Auto Ranging); Measures Voltage Tester, Current, Resistance, Continuity, Frequency; Tests Diodes, Temperature (Red)
WIDE RANGE of tests for AC/DC Voltage, AC/DC Current, Resistance, Continuity, Capacitance, Frequency and Temperature; Tests Diodes
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the actual news development?
Anthropic’s Institute has published a report arguing that AI is already accelerating AI development, supported by public benchmark trends and internal Anthropic measurements.
Has Anthropic shown that recursive self-improvement has started?
No. The report says recursive self-improvement has not arrived and is not guaranteed. Its evidence points to AI taking over more of the execution work, while humans still set goals and judge research quality.
What was the April 2026 agent experiment?
Anthropic says Claude agents worked on a weak-to-strong supervision problem, generated their own experiments and recovered 97% of a measured performance gap. Humans still chose the problem and scoring rubric.
Why does this matter outside AI labs?
If AI systems begin accelerating AI research at scale, the speed of model development could outpace current safety reviews, regulatory processes and institutional planning.
What remains unresolved?
The main open question is whether AI can automate research judgment, not only research execution. Anthropic’s evidence shows progress on the work of doing experiments, but less proof that AI can decide which experiments matter.
Source: Thorsten Meyer AI
