Claude Is Doing 26% of Anthropic’s R&D: Is AI Building AI?
A new disclosure from Anthropic, the company behind Claude AI, has put this shift into sharp focus. According to Anthropic's latest measurements, Claude now leads 26% of the company's AI research and development work as of August 2026.
That figure has grown dramatically from less than 1% earlier in 2026. At the same time, Anthropic says that more than 90% of its AI R&D work involves AI collaboration at some level.
But there is an important distinction:
Claude is not independently building its successor.
Humans are still involved, and Anthropic says Claude is not fully autonomous in any of the measured areas of AI R&D.
So what does the 26% figure actually mean?
And does this bring the AI industry closer to a future where AI systems can improve and build increasingly capable AI systems themselves?
Let's break it down.
What Happened at Anthropic?
Anthropic recently published a set of measurements designed to show how quickly AI is becoming involved in the development of advanced AI systems.
The company created an Anthropic R&D Automation Index to measure how much of its AI research and development work is being performed by Claude.
The measurement uses an automation scale developed by Epoch AI, ranging from:
AL0: No AI involvement
AL1: Minimal AI involvement
AL2: AI assists
AL3: AI collaborates
AL4: AI leads
AL5: AI operates autonomously
Anthropic reported that, as of August 2026, Claude was at the "leads" level for 26% of its AI R&D work.
However, Claude was not operating fully autonomously for any measured category of AI R&D.
This distinction is critical.
"AI leads" does not mean "AI works completely alone."
At the AL4 level, AI can complete most of a task from a high-level instruction while a human supervises the process.
How Fast Is AI's Role Growing?
The most interesting part of Anthropic's disclosure isn't only the 26% number.
It is the speed at which that number has increased.
Anthropic's data shows that Claude's share of AI R&D work at the "leads" level was under 1% in February 2026, before rising to 26% by August.
That represents a significant increase in just a few months.
More than 90% of Anthropic's AI R&D work was at or above the "collaborates" level, meaning AI was involved in a substantial way across the company's development process.
This suggests something important about the future of AI development:
AI is increasingly becoming part of the process used to create better AI.
That creates a potentially powerful feedback loop.
Better AI → better AI-assisted research → faster development → more capable AI.
Is Claude Actually Building the Next Version of Claude?
Partially — but not independently.
Claude is being used across many areas involved in AI research and engineering, including tasks related to model development, evaluation, infrastructure and software engineering.
Anthropic's methodology involved analysing approximately 15,000 granular AI R&D tasks from sampled employee work during July 2026. The company then organised those tasks into a hierarchy containing hundreds of categories and assessed how much AI was involved in each area.
This means Claude is becoming an increasingly important tool inside the process of creating future AI models.
But humans remain part of that loop.
Anthropic explicitly states that no measured subset of its AI R&D was fully autonomous as of August.
So the more accurate description is:
AI is helping humans build AI at an increasingly large scale.
What Does "AI Building AI" Actually Mean?
The phrase can sound dramatic, but the underlying concept is easier to understand.
Imagine an AI research team working on a new model.
Traditionally, humans might:
Design experiments
Write code
Analyse results
Identify problems
Test different approaches
Improve the training process
Evaluate the model
Fix infrastructure issues
Now AI systems can participate in many of these steps.
For example, an AI agent could:
Receive a task → analyse existing code → propose a solution → write code → run tests → analyse errors → make improvements → report the results to researchers.
A human can then review the work and decide what happens next.
The more capable the AI becomes, the larger the portion of this workflow it can handle.
That's the fundamental change Anthropic's data is showing.
Why This Matters for AI Development
There are several reasons this development is important.
1. AI Research Could Become Faster
If AI can perform more research and engineering tasks, human researchers may be able to run more experiments in less time.
Instead of researchers spending hours on repetitive engineering tasks, AI agents can handle parts of the workflow.
This could accelerate the development cycle of future AI models.
2. Smaller Teams Could Potentially Do More
AI-assisted development could increase the productivity of individual researchers and engineering teams.
A researcher who previously needed several tools and people to complete a task could increasingly delegate parts of the workflow to AI agents.
That could change how AI research organisations are structured.
3. AI Agents Are Moving Beyond Chatbots
Traditional chatbots primarily respond to prompts.
AI agents are different.
They can potentially:
Break a goal into smaller tasks
Use software tools
Write and execute code
Search information
Analyse data
Perform multi-step workflows
Monitor results
Iterate on their work
Anthropic's internal use of thousands of agents illustrates how AI is moving toward long-running, tool-using workflows rather than simple question-and-answer interactions.
Anthropic reported that approximately 30,000 agents were doing research and engineering work simultaneously on its most-used internal platform as of August 2026.
What About AI Safety?
This is where the story becomes even more important.
If AI systems are increasingly involved in developing more capable AI systems, human oversight becomes critical.
Anthropic says its internal agents are monitored using both online and offline systems.
According to the company's August measurements:
Around 30,000 agents were active on its most-used internal platform.
100% of their actions passed through online monitoring before execution.
More than 1 billion agent decisions were analysed during August.
About 0.002%, or approximately 1 in 47,000 decisions, were blocked by the online monitor.
Offline monitoring also reviewed agent activity after actions were taken.
Anthropic says the monitoring system is designed to identify potentially dangerous behaviour, including actions that could cause irreversible harm.
The company also says it plans to involve independent third-party evaluators to verify safety practices and monitor these metrics.
This matters because more capable AI agents require more capable oversight.
Does This Mean We Are Near AGI?
Not necessarily.
The 26% figure does not mean Claude has achieved AGI.
It also does not mean AI can independently replace AI researchers.
Anthropic's own measurements show that humans remain involved and that Claude is not fully autonomous in the measured R&D tasks.
However, the development is relevant to a much-discussed concept called recursive self-improvement.
What is recursive self-improvement?
Recursive self-improvement describes a hypothetical scenario in which an AI system can substantially improve its own capabilities, potentially use those improvements to create an even more capable system, and repeat the process.
The simplified concept looks like this:
AI → improves AI → better AI → improves AI again → increasingly capable AI
The 26% figure does not show that this loop is already happening autonomously.
Instead, it shows that AI is increasingly involved in the human-controlled process of developing future AI systems.
That distinction is extremely important.
Why Anthropic Published These Numbers
Anthropic says the purpose of these measurements is to give the public, governments and researchers greater visibility into the pace of AI development.
The company has proposed that AI developers regularly publish similar measurements so that progress can be tracked over time.
There are still challenges.
For example, different AI companies could measure their internal work differently, making direct comparisons difficult.
Anthropic also acknowledges that using AI models to evaluate AI work introduces methodological limitations.
For that reason, the company says independent verification could make future measurements more useful.
What Could This Mean for the Future of AI?
The implications go beyond Anthropic.
If similar trends occur across other frontier AI companies, AI could increasingly become part of the infrastructure used to develop future AI systems.
That could affect:
AI Research
Researchers could use AI agents for experiments, analysis, coding and evaluation.
Software Engineering
AI coding agents could handle increasingly complex engineering projects.
AI Startups
Smaller teams could potentially accomplish more with AI-assisted development.
Business Automation
Businesses may increasingly move from simple AI assistants toward autonomous or semi-autonomous agents.
AI Careers
The skills valued in AI-related jobs could shift toward:
AI system design
Agent orchestration
AI evaluation
AI safety
Data analysis
Prompt and workflow engineering
Human-AI collaboration
AI governance
The important lesson is not simply that "AI will replace humans."
Instead, the bigger shift may be:
Humans who know how to work effectively with AI could increasingly outperform workflows that rely only on traditional tools.
AI Is Becoming Part of the AI Development Loop
The biggest takeaway from Anthropic's announcement is not the number 26%.
It is the direction of change.
AI models are increasingly moving from being products used by researchers to becoming tools used to create the next generation of AI products.
That creates a new development loop:
Human researchers
↓
AI-assisted research
↓
More capable AI systems
↓
More powerful AI development tools
↓
Faster AI research
The loop is still human-supervised today.
But its growing importance explains why researchers, governments and AI companies are paying increasing attention to AI safety, transparency and oversight.
What Happens Next?
Anthropic's disclosure could become a benchmark for measuring how much AI is involved in frontier AI development.
If other major AI laboratories publish similar measurements, we may eventually have a clearer picture of:
How much AI is used to build AI
How quickly AI automation is increasing
How many AI agents are operating inside companies
How much compute is dedicated to AI research
How much human oversight remains necessary
Whether AI development is approaching more autonomous systems
For now, one thing is clear:
AI is no longer just helping us use technology. It is increasingly helping researchers build the technology itself.
And that could become one of the defining developments of the AI industry in the coming years.
Is Claude doing 26% of Anthropic's work?
No. Anthropic reported that Claude leads 26% of its AI research and development work as of August 2026. This refers specifically to AI R&D tasks, not all work performed by Anthropic.
Is Claude building itself?
Claude is increasingly involved in the research and development process used to build future AI systems, but Anthropic says it is not fully autonomous in any measured AI R&D category.
What does "AI leads" mean?
In Anthropic's measurement framework, "leads" means AI can complete most of a task from a high-level prompt while a human supervises it. It does not mean the AI operates completely independently.
How many AI agents does Anthropic use?
Anthropic reported approximately 30,000 agents doing research and engineering work at any one time on its most-used internal platform as of August 2026.
Is this recursive self-improvement?
Not yet in the fully autonomous sense. Anthropic's data shows increasing AI involvement in AI development, but the company says Claude was not fully autonomous in the measured R&D categories.
What is recursive self-improvement?
Recursive self-improvement is the idea that an AI system could improve its own capabilities and use those improvements to develop increasingly capable successors. It remains an important research and safety concept rather than something demonstrated by Anthropic's 26% figure.
Final Takeaway
Claude leading 26% of Anthropic's AI R&D is a major signal of how quickly AI-assisted development is evolving.
But the headline should not be interpreted as "AI has started building itself without humans."
The more accurate story is more interesting:
AI is becoming an increasingly important collaborator in the process of creating more advanced AI.
As models become better at coding, research, experimentation and long-running agentic tasks, the boundary between using AI and building AI with AI is becoming increasingly important.
For students, professionals, businesses and future AI builders, this means one thing:
Comments
Post a Comment