Anthropic says Claude now leads a quarter of its own AI research, up from 1% in February

From under 1% to 26% in seven months

The 26% number is part of a broader index Anthropic published on Wednesday to measure what share of AI research at the lab is actually being done by AI. The company built the "Anthropic R&D Automation Index" using a scale developed by Epoch AI, which grades each task from AL0 (no AI involvement) to AL5 (a fully autonomous AI). On that scale, Claude now leads 26% of the lab's AI research and development tasks, meaning it can carry most of a task from a high-level prompt while a human supervises.

That share was under 1% in February 2026. Above it, on more than 90% of research tasks, Claude is at or above the "collaborates" level defined as the model performing large portions of work under close human direction. Anthropic said no work was measured at the highest level of full autonomy, meaning every task still has at least one human in the loop.

Anthropic frames the metric as a way to track progress toward what it calls "recursive self improvement," or a model that can fully autonomously build its successor. CEO Dario Amodei has previously proposed a three step plan to slow AI development and warned about the technology's risk, while OpenAI disclosed last week that AI agents had hacked Hugging Face.

How Anthropic counted the work

To build the index, the company sampled 20% of its staff in July 2026 and broke their work into roughly 15,000 granular tasks, which the report organizes into 542 task nodes. Examples include "eval platform defect diagnosis and fixes" and similar research and engineering work.

A Claude based judge then assigned each task an automation level using only evidence from the same month or earlier. According to the report, the model agreed exactly with human raters on 59% of tasks, and within one level on 97%. Both numbers beat the 35% exact agreement rate among humans rating the same work.

On agents, Anthropic said about 30,000 agents run on its most used platform at any moment. Online monitors check 100% of actions before they execute, and of more than 1 billion decisions made in August 2026, 0.002% (about 1 in 47,000) were blocked. Offline monitors flag roughly 100,000 transcripts a week, with about 50 escalated as highest priority.

Where humans still drive the direction

Anthropic is careful to point out that the numbers describe work that humans were doing as recently as July 2026, and that the task basket it sampled misses new kinds of work that have emerged since. Between February and July 2026, the share of "novel" tasks did not rise, suggesting the model is absorbing existing work rather than expanding into new research areas on its own.

On compute, the company said roughly 6% of AI R&D compute went to safety work in the week of July 13 to 20, 2026, rising to 12% when looking only at AI driven AI R&D compute. The single week figures are meant to demonstrate feasibility rather than to establish a trend.

Anthropic says it will let independent third party evaluators access systems comparable to its internal risk teams, and plans to periodically rebuild the task basket so the index keeps reflecting current work.

What it means for AI jobs

For AI researchers and engineers, the report is the clearest public signal yet that the work of building frontier AI is itself being automated at the frontier AI labs. The same announcement comes against a broader shift in the AI labor market: Indeed Hiring Lab data published in July 2026 showed US software developer job postings on its platform had grown almost 15% since the launch of Claude Code in February 2025, reversing a multi year decline in postings for AI exposed roles.

The Anthropic numbers suggest that even at the labs where those tools are built, the demand for human labor on defined tasks is falling fast, while demand for engineers who can supervise, set direction, and evaluate AI research may rise. The company's index is unlikely to be the last attempt at the measurement. As Anthropic itself wrote, automation "could shift considerably if there were coordination on pacing the frontier," a sign that how fast AI takes over its own work is now a question that the labs, regulators, and workers all have reason to watch closely.