
Claude Now Leads 26% of Anthropic’s AI R&D: When AI Starts Developing AI, What Step Remains for Humanity?
Author: Robonaissance
Compiled by: TechFlow
TechFlow Read: For the first time, Anthropic used its own models to quantify the level of AI participation in its own R&D, finding that Claude has taken the lead on 26% of the work. This figure is both proof of technological advancement and a starting point for reexamining the role of humans. For practitioners, the key question is not whether AI can get things done, but where humans still stand.

Jack Clark Takes a Bet: AI Builds Itself Before 2028
On May 4, 2026, Jack Clark placed a bet in his Import AI newsletter for the day the “last human steps out of the loop.” He described the number as somewhat reluctantly given: By the end of 2028, the probability of an AI system appearing that is powerful enough to build its own successor without any human involvement is 60% or higher. For 2027, he gave 30%, explaining why he kept it low. AI research may require “creativity and heretical insight,” and so far, systems have not demonstrated this capability in any major way.
Five weeks later, on June 8, he wrote again: Over the past few months, he had been gathering evidence at Anthropic proving a smaller version of the same idea—the compound acceleration of the lab’s own productivity—has already begun. Then he asked his own question: “Do these amount to anything conclusive? No.”
On September 15, during an interview with NPR, Clark stated bluntly: “We do not have a third party to verify what we are saying.”
Two days later, on September 17, Anthropic released a statistic. The piece was written by Marina Favaro and Phillie Wright, with Clark listed as head of research direction. As of August 2026, according to the lab’s own figures, Claude led 26% of Anthropic’s AI R&D work, compared to less than 1% in February. Not a single measured type of R&D work ran without human involvement.
Which Steps Are Humans Still Standing On
The page on Anthropic gave Clark’s described event a name: recursive self-improvement—a model entirely autonomously building its successor. It treats this statistic as grounds to illustrate how close the world is to that moment. Beneath the number lies a more basic question: Who will be the last person remaining in the loop? What are they still doing?
Epoch AI’s hierarchy scale answers this with six steps, each defined by the human role. Figure 1 lists all six. Four of them display the pattern; the other two, unused and assistive, sit on lower steps.
At the “assistive” level, AI helps look up information and polish content. At the “collaborative” level, AI completes large chunks of work, but humans make decisions. At the “leading” level, AI completes most tasks based on a high-level directive, while humans handle oversight, correction, and approval. At the “autonomous” level, AI works completely independently, with no humans in the loop.
How fast progress is moving is another question. This statistic is Anthropic’s own measure targeting its own employees regarding the types of work they did in July 2026, evaluating how those tasks are typically completed, rather than what Claude is capable of doing.

The Epoch AI six-level automation scale applied by Anthropic, and the human role at each level. The proportions are Anthropic’s own statistics, as of August 2026, weighted by human hours in July for their AI R&D work. “Leading” is the fourth level; as of August 2026, the lab reported zero measured work at the fifth level, “fully autonomous.”
The Count Compiled by Claude Itself, Week by Human-Week
Who conducted the count? Claude did.
The inventory work behind that number was completed by Claude, not people. Each week in July 2026, one-fifth of employees were randomly selected from each department forming the model research loop. A Claude agent would go through each of their weeks. It read their Slack messages and internal documents, and recorded what work each person did.
Take one person and one week. That person receives one weight unit, distributed evenly across their various tasks. Four tasks get a quarter-unit each. Ten tasks get a tenth each. According to the page’s own description, this is a rough approximation, but performs reasonably well on average because task categories worked on by many accumulate the most weight.
Now add everyone in. Every person sampled each week in July 2026, after the same reading process, is aggregated into a flat list of approximately 15,000 fine-grained tasks.
These tasks are then placed into a tree. It contains 542 nodes of varying depths, 378 of which are leaves, representing the finest granularity. This tree is frozen, so monthly ratings are applied to the same basket of work. The leaves read like a week in a research lab: Evaluating platform defect diagnosis and remediation. Reinforcement learning sandbox egress and network policy. Service incident postmortem.
Each node then receives a rating. A Claude agent researches how this type of work is completed across the company—who does it, what tools are used, how much AI handles—it; an independent Claude reviewer reads the evidence and selects from the six levels. For any given month, the agent can access evidence from that month or earlier.
The page frames the “leading” level as what Claude can do, but the ratings follow the rules set by the Epoch scale authors: they describe how tasks are “typically completed” in frontier labs, not what could be achieved with maximum effort.
A hypothetical example on the lab’s page makes these steps concrete. A nighttime data pipeline breaks down and must be fixed before tomorrow’s run. At the “collaborative” level, an engineer passes logs to Claude, perhaps with an initial guess; when new issues arise, Claude pauses, and the engineer decides whether to bypass or fully fix them. After tests pass, the engineer might review the changes line by line, rerun the pipeline, and deploy it. At the “leading” level, the engineer simply feeds the fault alert to Claude, leaving the rest to it. It identifies the failing stage and cause, writes and tests a fix, reruns the pipeline on a data copy, compares the output to the last successful run, and documents what went wrong. Claude does not deploy. It flags the engineer, who reads the report, reviews the changes, asks a few questions perhaps, and then decides whether to ship tonight or hold off. At the higher, yet-unreached level, Claude would be trusted to monitor, fix, and deploy, while humans only intervene if they choose to.
How reliable are such ratings? Heads of each work area rate their own domains without seeing what the reviewers saw. The probability of agreement between reviewers and humans on specific levels is 59%, while human-to-human agreement sits at 35%. The probability that a reviewer differs from a human by no more than one level is 97%. The sample is not described, and weaknesses are plainly stated in the lab’s own words: “there is indeed room for divergence on edge cases,” such as where “collaboration” ends and “leadership” begins. A rating shifting from “collaborative” to “leading” would alter that 26% figure.
Figure 2 plots the lab’s monthly readings, rising from less than 1% in February to 26% in August, with monthly ratings applied to the same basket of July work. The lab also reported that work at the “collaborative” level and above accounts for over 90%.

This is Anthropic’s own rating, evaluated monthly by Claude assessors using the Epoch AI six-level automation scale against the lab’s July 2026 basket of model R&D work, weighted by July human-hours. The February value is self-described on the chart (under 1%); March through August reflect the labeled numbers. Anthropic’s chart also shows a 90% measurement interval and other automation tiers; not redrawn here. Leads represent the fourth tier of this scale, indicating human supervision; as of August 2026, the lab reported no measured work types reaching the fifth tier.
Thus, the composition of this number is: a week, a tree, an evaluator, and a boundary. The result is an aggregate total for the entire lab, not broken down by work type, making it impossible to specify which tasks fall within the 26%.
Epoch Rates Decision-Making Low, Code Writing High
In May, Clark drew a line between two types of work. He wrote that AI can now automate a large portion, perhaps all, of AI engineering, but it remains unclear how much of AI research it can automate.
Authors at Epoch similarly categorize lab work, with the first category being decision-making: what to do, what to continue, where to allocate compute. Epoch’s own sample rating (which they call a starting best guess) scores “deciding what to release” and “setting team direction” at 1, meaning marginal assistance, while scoring “writing research code” at 4, known as leads. Where the 26% falls between these two extremes is precisely the question this aggregate cannot answer.
Days described at the lab point in the same direction. According to an article co-authored by Clark, Claude delivered over 800 fixes in April, reducing a certain class of API errors by a thousandfold, while the supervising engineer estimated the workload would take a human four years to complete, and one employee hadn’t written code for roughly five months. Another employee wrote that when everything works smoothly, what they do seems insignificant; when everything crashes, “I’ve lost track of what I’ve even been doing.”
The lab’s August risk report gives a single-word reason for retaining manual steps: trust. Researchers could have deployed massive amounts of AI labor on valuable work, but they opted for a moderate volume because there are steps they do not trust the model to execute correctly. The report draws a second conclusion from this same fact. The lab had ample reason to replace researchers with models, but it did not. The report interprets this as a signal that models still cannot replace an entire research team. It also states this interpretation has only been partially tested: the lab has never spent five times the cost of researcher resources on tasks where models failed, to verify whether money can bridge the gap.
The situation at OpenAI is similar. As of mid-August, its research organization runs 3.1 agent-workdays per human-workday, with users in the 90th percentile spending over $7,000 daily on tokens. People there still set research priorities, judge which ideas and results are worth pursuing, and decide whether to scale up, pause, or deploy.
Engineering is the domain Clark has already handed over to machines. Research remains the question he left for later.
The Lab’s Own Rules Look at Speed, Not Proportion
The June statement is about speed. The compound acceleration of lab productivity is a statement about the rate of progress, whereas this number measures something else: the share of work handed off.
On September 30, Chris Painter, president of the external AI model evaluation agency Model Evaluation and Threat Research (METR), entered the 26% into the Congressional record as written testimony for a hearing titled “Rogue AI: Securing the Homeland Against AI Agent Attacks.” In subsequent text, he wrote that when companies use AI to automate AI research, feedback loops “could lead to self-sustaining acceleration of capabilities.”
What triggers this threat under Anthropic’s own rules is speed, not proportion: “rapid acceleration,” defined as doubling the rate of progress, meaning gains achieved in one year equal baseline two years. The rule states this does not equal a doubling of researcher productivity, because doubling inputs does not necessarily double the rate of progress.
According to the interpretation in the August risk report, during the period ending July 15, the lab’s research speed was significantly faster than without AI assistance, but had not yet doubled. The lab is uncertain about this, and measuring it is difficult. Early signs of acceleration have appeared, and the lab may cross the threshold within the next year. Five days after Anthropic published the data on September 17, a system card for Claude Opus 5.5 noted that internal lab measurements showed no sustained, attributable two-fold AI-driven acceleration.
METR also released a summary of its pre-release assessment for Claude Opus 5.5 on September 22. Within it, METR paraphrased an estimate from another METR team that has greater access than the summary authors: overall capability acceleration driven by AI is approximately 1.5x, with about a 30% probability of reaching 2x. The team could not share its evidence with the summary authors, and the estimate does not specify the covered time period.
The September 17 number measures the degree of handoff. The August report and system card estimate the rate. None of the three simultaneously performed both measurements on the same type of work within the same time frame.
Interpreting 26%
By the lab’s own accounting, as of August 2026, Claude leads 26% of Anthropic’s AI research work, and this number has risen for six consecutive months on a basket sample of July work. This is a handoff calculated per the lab’s metrics, collected by Claude and cross-referenced with ratings from employees responsible for the work.
This cannot yet be interpreted as how close a closed loop is. The lab’s rules focus on speed, and the lab’s own estimates fall below the line drawn by those rules. The unresolved question this number leaves is what falls inside that 26%. If the 26% is primarily engineering, it measures the part Clark already considered within reach in May. If it touches decision-making and research, it measures the part he left for later.
The 26% points to engineering. Ratings place decision-making at a lower tier, and the lab’s examples leave release decisions to humans, with researchers still maintaining trust. This is an interpretation, not a measurement, and it could be verified with data split by work type. Decision-making at lower tiers and research code at the leads tier would confirm this interpretation. Decision-making reaching the leads tier would push the 26% toward the research end.
Among the descriptions of the highest tier on the page, the clause regarding humans is: full intervention is unnecessary unless humans choose to participate themselves. One tier below that, according to the lab’s own example, the last person is an engineer holding a report, deciding whether to ship tonight.
Reading the frontier. Close readings of what frontier labs publish, unfolding through frameworks that explain their significance.













