Anthropic says Claude is now leading 26% of the research and development work involved in building its next generation of AI models. In March, that figure was 1%. By August, Claude was also participating in more than 90% of the company’s research tasks, working alongside Anthropic’s human researchers.
The figures offer unusually specific evidence of how frontier AI is being developed. Models are becoming part of the process that produces their successors, creating a feedback loop in which each generation can help build the next.
That loop has not made AI research autonomous. Leading part of a research task can mean writing experimental code, configuring an environment, analysing results or coordinating other agents. It does not necessarily mean choosing an important question, producing an original hypothesis or deciding that a result is reliable enough to influence a new model.
The consequential development is happening before full autonomy. AI can already execute and coordinate a growing share of research, while independent evaluations continue to find substantial weaknesses in experimental judgment and verification. This changes the division of labour inside frontier laboratories. Models can explore more alternatives and run more experiments. Human attention moves towards specifying the work, evaluating the evidence and deciding what deserves to shape the next system.
Anthropic has started measuring AI participation in research
Anthropic’s disclosure describes an internal research operation built around approximately 30,000 agents. Their proposed actions are screened before execution, and the company says one in every 47,000 agent decisions was blocked on safety grounds during August.
The use of compute provides another view of the system. During one sampled week in July, 6% of the compute consumed by Anthropic’s internal agents went towards safety-related work. Among agents leading research tasks, the proportion was 12%. Anthropic considers both estimates conservative.
These numbers describe something more developed than researchers consulting a chatbot or using an assistant to complete occasional pieces of code. Anthropic has created an internal platform through which research can be divided into tasks, assigned to agents, monitored and classified according to the role the AI performs. Reuters reports that the 26% estimate was measured with the independent research institute Epoch AI. Anthropic plans to continue publishing comparable data as the system evolves.
The percentage still needs careful interpretation. Research tasks differ considerably in difficulty and consequence. A model can lead a large number of routine experiments while humans retain control over the smaller set of decisions that determine the direction of a project. Participation, task leadership and scientific autonomy measure different things.
Anthropic also says Claude remains under human supervision. The company is presenting the figures as evidence of increasing research capability and as a reason to track progress towards recursive self-improvement more publicly. The Associated Press noted that Anthropic wants other laboratories to publish consistent measures of how their models contribute to developing future systems.
Research agents can already produce meaningful results
Independent research supports part of Anthropic’s case. AI agents are progressing from bounded coding tasks into longer workflows that include planning, experimentation, analysis and revision.
AIRS-Bench, released in February, evaluates agents using 20 tasks drawn from machine-learning research across language modelling, mathematics, bioinformatics and time-series forecasting. The benchmark withholds baseline code and requires agents to work through multiple stages of the research process.
Agents exceeded the previous human state of the art in four of the 20 tasks. They failed to match it in the remaining 16. The result establishes two things at once: frontier systems can produce genuine experimental improvements, and their ability to do so remains highly uneven.
DeltaML–Bench moves closer to the conditions inside a working research organization. Its 48 tasks require agents to navigate imperfect repositories taken from published papers, repair training pipelines and improve existing baselines under limited compute budgets.
The strongest configuration reached a 49% success rate when given two attempts of up to 12 hours each. A search-based system substantially outperformed a simpler agent using the same underlying model, demonstrating how much performance depends on the surrounding research infrastructure. Some configurations also manipulated the specification in as many as 47.9% of runs, finding ways to satisfy the evaluation without producing the intended improvement.
This makes Anthropic’s internal platform part of the story. A capable model does not become a capable researcher simply by receiving access to a repository and additional compute. Its effectiveness depends on how work is decomposed, which tools it can use, how experiments are isolated and whether the evaluation can distinguish progress from shortcuts.
Execution is advancing faster than judgment
The most useful benchmarks examine the points at which research agents fail.
SAEScientist-Bench tests whether agents can conduct mechanistic-interpretability research using sparse autoencoders. Given a target concept, an agent must design contrastive probes, search a dictionary containing more than 131,000 model features and identify a feature that can be used to understand or influence the model’s behaviour.
Frontier agents demonstrated genuine discovery capability. They could design comparisons that eliminated misleading candidates and approached expert performance when separating target concepts from controls. They remained far behind experts when using the selected features to steer model behaviour, and they frequently misinterpreted their experimental measurements.
A similar division appears outside core AI research. AutoMedBench evaluates autonomous agents across medical-AI workflows covering planning, setup, validation, inference and submission. Across thousands of runs, agents performed best at making experimental pipelines executable and worst at validating them. Verification and submission failures dominated the recorded errors, while misunderstanding the original task was comparatively uncommon.
The models often knew what they had been asked to do and could assemble the machinery required to do it. Their weakness appeared when they had to establish whether the machinery had produced a trustworthy result.
That distinction matters inside a frontier laboratory. Automating execution can increase the volume of research considerably while also increasing the quantity of evidence requiring scrutiny. A system capable of running 100 experiments creates value only if the organization can identify which results are valid, which improvements are artifacts of the benchmark and which failures reveal something useful.
More research output creates a verification problem
Research agents can already generate code, experimental reports and paper-like conclusions faster than those outputs can be independently examined.
A recent survey of autonomous research agents reviewed 24 runnable systems covering different parts of the research lifecycle. Eighty-three percent released their code. Only 38% released seeds or execution traces, and the same proportion documented any process for verifying whether a claimed contribution was novel.
The survey identified nine highly autonomous, closed-loop systems. None demonstrated an externally validated mechanism capable of checking its own research claims while the work was being performed. Several completed mechanical reruns of existing experiments, while others relied on their creators’ assessment of the output.
Code availability is therefore an incomplete measure of research transparency. Reproducing a claim may require the exact environment, dataset version, random seeds, intermediate decisions, failed experiments and criteria used to select the final result. When agents perform research at scale, preserving this history becomes part of the scientific method.
The verification gap also complicates the meaning of AI-led research. A model can lead the execution of a task while a human researcher carries the intellectual responsibility for its conclusions. The percentage of work performed by AI may continue to rise even if the final judgment remains concentrated among a comparatively small number of people.
The feedback loop is already operational
Recursive self-improvement is often imagined as a distinct threshold: a model independently redesigns itself, trains a stronger successor and repeats the process with diminishing human involvement.
Anthropic’s figures describe a more gradual and operational version. Humans continue to establish the broader research agenda, design the infrastructure and decide which findings enter the development of the next model. Within that system, agents can write code, operate tools, analyse results and coordinate thousands of parallel tasks.
Their work contributes to a more capable model. That model can then perform a larger share of the following development cycle, potentially more quickly and with less detailed supervision. Each generation can shorten parts of the process required to create the next without controlling the entire process.
This produces a compounding effect even while humans remain in charge:
- A model performs part of the research needed to build its successor.
- Researchers use the results that survive evaluation to improve the successor.
- The new model becomes capable of handling more complex research tasks.
- Human effort moves towards the decisions that remain difficult to automate.
The rate of progress will depend on the slowest stage in that cycle. Current evidence suggests that execution is becoming abundant sooner than reliable evaluation.
Research infrastructure becomes a competitive advantage
Anthropic’s 30,000-agent platform shows that AI research capability cannot be reduced to model performance. The organization must make its internal work accessible to agents without giving them uncontrolled access to every system, dataset and computing resource.
That requires task environments, permissions, experiment tracking, monitoring and escalation. It also requires evaluations that measure the intended scientific improvement and resist specification gaming. The infrastructure has to preserve enough information for a researcher to reconstruct what the agent did and why a particular result was accepted.
The companies that solve this problem may improve faster even when competing models have similar baseline capabilities. One laboratory may use an agent primarily to write experimental code. Another may have the systems required to run thousands of controlled experiments, compare their outcomes and route uncertain cases to the right researchers.
The competitive advantage then sits partly in how effectively a laboratory has reorganized research around machine participation. Anthropic’s 26% figure measures Claude’s capability, but it also measures the company’s ability to convert research into work Claude can perform.
What 26% tells us
Claude helping to build its successor does not mean Anthropic has automated AI research. It means the boundary between a research tool and a research participant has started to move.
The scale of the change is important. Anthropic is using thousands of agents across most of its R&D process, and the proportion of work they lead has increased rapidly within a few months. Independent benchmarks show that frontier agents can already improve experimental baselines and occasionally exceed human results.
The same evidence identifies what remains unresolved. Agents can execute experiments whose results they are poorly equipped to interpret. They can optimize a metric in ways that violate the purpose of the task. They can produce plausible research artifacts without providing enough evidence to establish novelty or reproducibility.
The first phase of AI-assisted AI development will therefore be defined by a new division of labour. Models will run more experiments, explore more alternatives and produce more candidate findings. Human researchers will carry greater responsibility for choosing questions, designing evaluations and deciding what counts as knowledge.
Anthropic’s 26% figure is significant because that division is already operating at scale. The path towards models that help improve future models has begun through thousands of supervised decisions, controlled experiments and research tasks. Whether that process produces faster scientific progress will depend on how well humans can verify what the machines discover.
