📊 Full opportunity report: When AI Builds Itself: Inside Anthropic’s Evidence on Recursive Self-Improvement on ThorstenMeyerAI.com — validation score, market gap, and execution plan.

TL;DR

Anthropic reports that AI models like Claude are rapidly automating tasks involved in AI research and development. Evidence shows significant progress in AI-generated code and experiments, but autonomous goal-setting remains unachieved. This raises the possibility of AI self-improvement loops forming sooner than expected.

Anthropic has published new internal data indicating that AI models like Claude are increasingly capable of automating core tasks in AI research and development, potentially enabling self-improving AI systems. This development is significant because it suggests that, under certain conditions, AI could begin improving itself at a rate faster than human-led efforts, a possibility that has long been debated but lacked concrete recent evidence.

The report from The Anthropic Institute details how AI models are now capable of performing a majority of coding and experimental tasks involved in AI development. For example, as of May 2026, over 80% of code integrated into Anthropic’s projects was authored by Claude, up from single digits in early 2025. Public benchmarks like METR, SWE-bench, and CORE-Bench show rapid improvements in AI capabilities, with models now handling tasks that previously required days of human effort, often within hours or less.

Inside the labs, data indicates that AI can already take on roles traditionally performed by less experienced researchers, such as fixing bugs or reproducing scientific results. While the models are improving at executing specified tasks, the report emphasizes that the critical gap remains in AI’s ability to set its own research goals, a key step toward autonomous self-improvement. The authors stress that this gap is the primary obstacle preventing current AI from fully self-optimizing its development process.

When AI builds itself — ThorstenMeyerAI.com
ThorstenMeyerAI.com
The Anthropic Institute · Deep-Dive
recursive self-improvement · the evidence

Table of Contents

When AI builds itself

Anthropic is delegating a growing share of AI development to AI. Taken far enough, that points to a system that designs its own successor — recursive self-improvement. Not here yet, not inevitable. But the case isn’t speculation: it’s data on what AI is doing to AI development right now.

8× code/engineer · >80% of merged code by Claude · benchmarks saturating · the human role narrowing
AI can increasingly do the doing of AI research — writing code, running experiments, producing results. Humans still hold the deciding — which problems matter, which results to trust, when an approach is dead.
Recursive self-improvement is what happens if that last human-held piece — research taste — also falls to automation. Every result below is a rung on the ladder from “the doing” toward “the deciding.”
01Evidence from outside

The curve that hasn’t bent

METR tracks the length of tasks AI can reliably complete on its own. That horizon is doubling roughly every four months — up from every seven. Anyone can check this in public data.

Task horizon — how long a job AI can handle solo

Each model handles dramatically longer tasks than the one a year before. The line keeps going up.

Claude Opus 3
Mar 2024
~4 min
Claude Sonnet 3.7
~Mar 2025
~1.5 hours
Claude Opus 4.6
~Mar 2026
~12 hours
Claude Mythos Preview
2026
“at least” 16 hours
If the trend holds: tasks that take a skilled person days come into range this year; week-long tasks in 2027. (Mythos is already at the upper edge of what METR can measure without harder tasks.)
SWE-bench · real bug fixes
Low single digits → saturated in two years.
CORE-Bench · reproducing papers
~20% (2024) → saturated 15 months later. A prerequisite for original research.
02The framework
Coding with AI For Dummies (For Dummies: Learning Made Easy)

Coding with AI For Dummies (For Dummies: Learning Made Easy)

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Two kinds of work, one persistent gap

Building a frontier model splits into engineering and research. Across both, the pattern is the same — and so is the one thing AI still can’t do well.

engineering

Code, infrastructure, training

Claude can take an underspecified problem and find a method. Humans supply the goal; they no longer need to supply the method.

✓ method: solvedgoal-setting: gap
research

Which experiments, what they mean

Claude can match or outperform skilled humans at executing a well-specified experiment. But choosing which experiment still needs a human.

✓ execution: strongtaste: gap

The same ladder Anthropic employees climb with experience

junior
Execute a set task: “The export button isn’t working, please fix it.”
experienced
Design the approach: “Investigate why the network slows down under heavy load.”
senior
Choose what’s worth doing: “What should the team build next quarter?”
03The narrowing role · step through it
CLAUDE AI UNLEASHED From First Prompts to Pro: The Complete Guide to Claude AI for Writing, Research, Coding, and Business (The Claude AI Mastery Series)

CLAUDE AI UNLEASHED From First Prompts to Pro: The Complete Guide to Claude AI for Writing, Research, Coding, and Business (The Claude AI Mastery Series)

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Watch the human share shrink, rung by rung

Walk up the four stages of AI development. At each, the human/AI split shifts — and the real internal numbers show exactly where AI has reached parity, gone superhuman, or still trails. Tap a rung.

The human role across the development loop

The doing now costs almost nothing in human time. What’s left is the deciding.

⌨️
Write code
⚙️
Run experiments
💡
Propose experiments
🧭
Set direction
the doingthe deciding
AI does this human does this
04The headline result
Agentic Development: The Complete Guide to AI-Assisted Coding with Claude, Cursor, and Beyond

Agentic Development: The Complete Guide to AI-Assisted Coding with Claude, Cursor, and Beyond

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Agents ran an open research project end to end

April 2026: the first demonstration of Claude running an open-ended research project from hypotheses to findings — on a real AI-safety problem.

weak-to-strong supervision

Can a weaker model reliably supervise a stronger one?

Agents were left to solve it: proposing hypotheses, testing them, sharing findings across parallel agents, iterating. Measured against the gap between a “floor” (weak supervisor alone) and “ceiling” (strong model trained on correct answers).

share of the floor→ceiling gap recovered
agents: 97%
humans: 23%
97%
recovered by agents
(humans: ~23% in a week)
800 hrs
cumulative agent time
· ~$18,000 compute
every one
experiment designed by
the agents themselves
The caveats are load-bearing — and Anthropic states them: the result didn’t transfer cleanly to production-scale models, and humans still chose the problem and wrote the scoring rubric. The agents were superb inside the frame. The frame was still human. That boundary is the whole story.
05The first climb toward taste
Architecting Data and Machine Learning Platforms: Enable Analytics and AI-Driven Innovation in the Cloud

Architecting Data and Machine Learning Platforms: Enable Analytics and AI-Driven Innovation in the Cloud

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Picking a better next step than the human

Real research sessions where a human took a wrong turn. Models saw only the work before the detour and proposed a next step; a judge that knew the outcome scored them. The day-to-day of research is this chain of next-step calls.

“Can the model pick a better next step than the human?”

Share of moments where the model’s next step was judged better. The amber line is the practical ceiling (an ideal answer that could see the whole session).

Opus 4.5
Nov 2025
51%
Mythos Preview
Apr 2026
64%
Read this carefully — Anthropic insists on the asterisk: these n=129 moments were deliberately chosen because the human’s choice had room for improvement, so it’s not a like-for-like human-vs-model comparison. On a separate set where the human’s move was already strong, models won only ~20% of the time. The honest reading: where a human stumbled, AI increasingly offered the better recovery — and that’s rising.
06Three futures, held honestly

It depends on whether the trend continues — and what we do

The piece refuses a single prediction. It lays out three scenarios, and is clear about which it finds most likely.

1
the trend stalls, capabilities diffuse

The exponentials turn out to be S-curves

Maybe taste can’t be scaled into existence; maybe the constraint is the supply chain — chips, grid, interconnect — not intelligence. Even so, the world still changes: Glasswing’s Mythos found 10,000+ critical vulnerabilities in weeks, and a 100-person firm does the work of 1,000.

included for completeness · they doubt it
2
compounding efficiency gains

Development automates; humans still steer

100-person companies doing the work of tens of thousands — revolutionary, but turnable to harm (population-scale surveillance, tailored manipulation). Bound by Amdahl’s law: speeding one part shifts the bottleneck — which is exactly why human code review became Anthropic’s new chokepoint.

★ they think we’re likely heading here
3
full recursive self-improvement

AI designs and refines its own successors

Progress paced only by compute. Humans move to oversight of an expanding “virtual lab.” The future they understand least — especially whether alignment holds, or whether rare misalignments compound as models build successors, until control slips.

the one they’re most uncertain about
07The ask · & reading it straight

Build the option to slow down — verifiably

The piece ends on policy, not product. A unilateral pause just changes who leads; what’s missing is the ability to verify others have actually slowed.

Why a credible pause is hard — and worth building toward

A slowdown that only lets the least cautious catch up leaves everyone less safe. So the goal is the option: systems that let frontier labs verify others have genuinely stopped. Anthropic says if such systems existed and peers paused verifiably, it expects it would too.

why it’s hard
Detection beats verification — and even that’s tough

Training runs are easier to conceal than missile silos, inputs are general-purpose, and whoever continues while others pause inherits the lead.

the precedent
We’ve done it before — slowly

Regimes like the INF Treaty built verification and trust over decades. The authors’ blunt line: “We don’t have that long.”

Reading it in proportion

  • This is one lab’s account of its own internal data — much previously unreported, not independently audited.
  • The soft spots are stated in the original: lines-of-code overstates productivity; the self-reported 4× is probably high; the headline research result didn’t transfer to production scale; the next-step test used cherry-picked moments.
  • “More autonomous” is not “fully autonomous” — every standout result still had a human framing the problem and defining success.
  • That the authors surface these caveats themselves — against their own incentive — is part of what makes the document serious.
ThorstenMeyerAI.com
Source: “When AI builds itself,” Marina Favaro & Jack Clark, The Anthropic Institute · data via METR, SWE-bench, CORE-Bench & Anthropic’s published research · figures per the piece · independent commentary.

Implications of AI Accelerating Its Own Development

This evidence suggests that AI is already significantly automating parts of its research cycle, which could lead to faster innovation and potentially autonomous self-improvement if the ability to set goals and priorities also automates. The possibility raises questions about future control, safety, and the pace of AI progress, making it a critical development for researchers, policymakers, and industry stakeholders.

Recent Trends in AI Self-Development Capabilities

Over the past several years, AI models have shown steady improvements on benchmarks measuring their ability to perform coding, scientific reasoning, and task automation. Public data from METR and other benchmarks reveal a doubling of AI capability horizons approximately every four months, a significant acceleration from previous rates. Inside labs like Anthropic, data indicates that models are increasingly capable of automating tasks that support AI research, such as code generation and experiment reproduction, with the most recent figures showing a dramatic rise in AI-authored code and experimental outputs.

While these trends are clear, the internal data also reveals persistent gaps in AI’s ability to autonomously determine which research problems to pursue, a necessary step for true recursive self-improvement. Experts caution that despite rapid progress, full self-directed development remains a future milestone rather than an imminent reality.

“The data from Anthropic strongly suggests that AI is already automating significant portions of the research process, which could accelerate development cycles considerably.”

— Thorsten Meyer, AI researcher

Unresolved Questions About Autonomous Goal-Setting

It is still unclear whether AI systems will soon be able to autonomously set and pursue their own research goals without human input. The report emphasizes that current models are strong at task execution but weak at goal formulation, which remains a significant barrier. The timeline for overcoming this gap, and whether it will happen naturally or require new breakthroughs, is still unknown.

Next Steps for Monitoring AI Self-Improvement Progress

Researchers and industry leaders will likely focus on developing benchmarks and internal metrics to better measure AI’s ability to autonomously define research objectives. Further transparency from labs about internal data and capabilities will be crucial. Additionally, safety and control mechanisms will need to evolve in tandem with these technological advances to mitigate potential risks of rapid self-improvement.

Key Questions

What does recursive self-improvement mean in AI?

Recursive self-improvement refers to AI systems that can improve their own design, code, or capabilities without human intervention, potentially leading to rapid, exponential progress.

How close are we to AI that can self-improve autonomously?

While current models are automating many research tasks, the critical ability to set their own goals remains unachieved. Experts believe full autonomous self-improvement is still a future milestone, not an imminent reality.

What are the risks of AI self-improvement?

Potential risks include loss of human oversight, unpredictable behavior, and rapid escalation of capabilities. Ensuring safety and control measures are in place is a key concern for researchers and policymakers.

What role do benchmarks play in assessing AI progress?

Benchmarks like METR, SWE-bench, and CORE-Bench measure AI performance on specific tasks, providing data on capability growth but not on internal development pace or goal-setting ability.

Will AI self-improvement lead to superintelligence?

This is uncertain. While rapid self-improvement could accelerate AI capabilities, whether it will result in superintelligence depends on many technical and safety factors that remain unresolved.

Source: ThorstenMeyerAI.com

You May Also Like

The Twelve Real Complaints About AI Tools in 2026 — A Reddit, Twitter, and GitHub Synthesis

A detailed report on the twelve most common issues users report with AI tools in 2026, highlighting reliability, capacity, and trust concerns.

Mobile Banking: How Your Phone Became Your Bank

How has mobile banking transformed financial management? Discover the revolutionary features that make your smartphone an essential banking tool.

The Continual Learning Research Map: Where the Memento Constraint Stands in May 2026

Six months on, the research community confirms the Memento constraint as a key bottleneck in frontier AI continual learning, with no ready solutions yet.

Forward-Deployed Engineer Economics 2.0: The Unit Economics Math, Six Months Later

Six months after initial analysis, FDE unit economics reveal profitability at enterprise scale but risks at lower levels, influencing AI lab scaling strategies.