Recursive Self-Improvement Explained: How Claude Now Helps Build the Next AI

Quick answer: recursive self-improvement is the idea that an AI system could help design, build and improve its own successor, with less and less human input each time. It is no longer just theory. In September 2026, Anthropic said its Claude models now lead about 26% of its research and development work, up from almost none in February 2026, and collaborate on roughly 90% of it. Humans still supervise and choose the research direction, so AI is not yet building itself on its own.

This guide explains what recursive self-improvement means in simple terms, what Anthropic’s numbers actually show, the caveats Anthropic itself gave, how other labs compare, why it matters for safety, and what it means for people who use AI tools every day.

What Is Recursive Self-Improvement?

Think of it as a loop:

1. Engineers build an AI model.
2. That model helps write code, run experiments and analyze results for the next model.
3. The next model is better, so it helps even more with the model after that.
4. Each cycle could get faster, with humans doing less of the work.

“Recursive” means the process repeats on itself. At the extreme, AI would improve AI with almost no human help. That end point has not been reached, but Anthropic’s report suggests the early stages of this loop are already happening inside AI labs.

A simple analogy

Imagine a factory that builds machine tools. At first, people do all the work. Then the factory’s own machines start helping to build the next, better machines. Each new generation of machines makes the next one faster and cheaper to build. People still run the factory and decide what to make, but the machines are doing more and more of the actual building. That is roughly where AI labs are today with their own models.

What Anthropic’s Report Shows

Anthropic published its findings in a report titled “When AI Builds Itself”, together with an announcement covered by news outlets across the USA on September 17 and 18, 2026. The key numbers:

“Leading” a task means Claude completes most of it from start to finish based on a high-level instruction, while a human supervises. “Collaboration” means Claude does large parts of the work under close human direction.

How Fast Claude’s Abilities Have Grown

One of the clearest ways to see the trend is the length of task a model can handle on its own:

That is a jump from minutes to a full working day in two years. The independent research group METR, which runs this kind of benchmark, also found that one of Anthropic’s most advanced research models could work for at least 16 hours.

The Experiment That Stood Out

Anthropic described a research test on an optimization problem. Two human researchers working for about a week recovered roughly 23% of the target performance gap. Claude agents, working for about 800 cumulative hours with roughly $18,000 | AED 66,105 in computing costs, recovered about 97%.

It is an impressive result, but Anthropic was clear about its limits: the gains did not transfer cleanly to full production-scale models, and humans still chose the problem and wrote the scoring system used to judge success.

It also shows why Anthropic tracks cost alongside speed. Running hundreds of agent hours is not free, so labs have to weigh how much computing to spend on automated research versus training and serving models for customers.

The Caveats Anthropic Gave

The report is unusually honest about where the numbers fall short:

• Lines of code are not productivity: Anthropic says the 8x code figure is almost certainly an overstatement of the true productivity gain.
• Code quality took time: Claude-written code was noticeably weaker than human code until recently and has only reached rough parity.
• The code share is conservative: the 80% figure counts only code clearly attributed to Claude, so the real share may be higher.
• Judgment still lags: models remain behind humans at deciding which research directions are worth pursuing, even if they execute quickly once a direction is chosen.
• Humans stay in charge: oversight measures remain in place to catch mistakes or misbehavior by the agent workforce.

How Other AI Labs Compare

Anthropic is not the only lab automating its own research. OpenAI reported that by mid-August 2026 its research teams ran 3.1 agent-workdays of coding-agent effort for every human workday, which we explain in our AI coding agents guide.

The two companies measure different things, so the numbers are not directly comparable. But both point the same way: AI is doing a growing share of the hands-on work of building AI.

Why This Matters for AI Safety

Recursive self-improvement is one of the central topics in AI safety. If AI speeds up its own development, new systems could arrive faster than people can test and understand them. That concern was at the heart of former Anthropic researcher Jacob Coxon’s public resignation, which we cover in our AI safety explainer.

Anthropic argues that publishing these numbers is part of the answer. Its report says labs should do everything possible to minimize the gap between what they know and what the public knows. Anthropic CEO Dario Amodei has supported the idea of a verifiable, industry-wide slowdown, and the Anthropic Institute says it will research verification systems that could make such a slowdown possible. Others in the industry disagree and say companies can manage the risks themselves.

Three Ways This Could Play Out

Experts disagree about where this trend leads. The main views fall into three broad camps:

• Steady progress: AI keeps speeding up research, but human judgment, computing power and energy limits keep the pace manageable. Labs and regulators have time to adapt.
• Rapid acceleration: AI takes on more of the research loop, including choosing directions, and progress speeds up sharply. This is the scenario safety researchers worry about most.
• Diminishing returns: automation helps with routine engineering, but breakthroughs still depend on human insight, so the loop slows down.

Anthropic’s own data shows strong acceleration in execution, while judgment still lags. Which path the industry follows will depend on technical progress, business pressure and the rules governments set.

What It Means for Jobs

The same tools that help AI labs are reaching ordinary workplaces. Software engineers, data analysts and researchers in the USA and UAE are already using AI agents for routine tasks such as writing tests, cleaning data and drafting reports. The skills growing in value are the ones AI still struggles with: choosing the right problem, judging quality and taking responsibility for decisions.

What It Means for Claude Users

• Faster improvements: if AI helps build AI, the tools you use, from Claude to ChatGPT, are likely to keep improving quickly.
• Longer tasks: as models handle longer tasks on their own, features like background tasks and AI agents become more practical. Our Claude Docs, Slides and Design guide shows what that looks like today.
• No change to your plan: this is a report about how Anthropic works internally, not a new feature. Claude’s Free plan and Pro plan at $20 | AED 73 a month are unaffected.
• Keep checking results: even very capable AI makes mistakes, so verify important facts and code.

The Business Angle

The report also matters for investors. Anthropic is preparing for a possible stock market listing, explained in our Anthropic IPO guide. Showing that its research is speeding up, while being open about the limits, helps explain why investors value AI labs so highly and why safety questions matter to them too.

How to Follow This Topic

• Company reports: Anthropic and OpenAI now publish data on how they use their own models. Read the original reports, including the caveats, not just headlines.
• Independent testers: groups such as METR measure model abilities using the same tests over time, which makes trends easier to compare.
• Policy news: watch for new AI rules in the USA, the UAE and Europe, especially around testing and transparency for the most powerful models.

Key Terms Explained

New to these ideas? Start with our explainers on what an AI agent is and Claude AI explained.

Frequently Asked Questions

What is recursive self-improvement in AI?

A scenario where an AI system helps design, build and improve its own successor, with each cycle needing less human input. It has not fully happened yet, but AI labs report early steps toward it.

Is Claude building itself without humans?

No. Anthropic says humans supervise throughout and still choose which research directions to pursue. Claude leads about 26% of R&D tasks from a high-level instruction, under human supervision.

How much of Anthropic’s code does Claude write?

More than 80% of the code merged into Anthropic’s production systems as of May 2026 was written by Claude, according to the company.

How many Claude agents does Anthropic run?

Roughly 30,000 Claude agents were doing research and engineering work at the same time as of August 2026.

Does this change Claude for regular users?

Not directly. It is a report on Anthropic’s internal work, not a new feature or price change.

Why is recursive self-improvement a safety concern?

If AI speeds up its own development, new and more powerful systems could arrive faster than people can test and understand them. That is why some researchers call for more transparency and outside checks.

Image credits: Photo: TechCrunch, Wikimedia Commons, CC BY 2.0.