Anthropic Claude Self-Improvement Is Accelerating, but Humans Still Choose the Destination

Anthropic says Claude now leads 26% of its model research and development, despite leading none of that work in February 2026. That sharp rise makes Anthropic Claude self-improvement more than a provocative slogan. The company says its model can complete large assignments from high-level prompts, although humans still supervise the work.

The disclosure pushes a familiar coding trend into more consequential territory. Claude is not merely helping engineers build applications. It is increasingly writing infrastructure, running experiments, analyzing results, and reviewing changes used to develop future Claude models.

Anthropic calls this progress toward recursive self-improvement, where an AI system contributes to building a more capable successor. Yet its evidence also exposes the remaining barrier. Claude executes defined work quickly, while people still decide which problems matter and whether the results deserve trust.

That distinction places Anthropic’s claim between two competing interpretations. One describes a compounding development loop that accelerates with every model generation. The other describes advanced automation operating inside a research agenda that humans still control.

Anthropic Claude Self-Improvement Now Reaches Model R&D

The important change is not that Claude writes code. It is that Claude now leads substantial parts of the work behind its successor.

Anthropic disclosed the 26% figure on September 17, 2026. The company defines “leading” as completing most of a task from a high-level prompt while remaining under human supervision. Claude led none of the measured research and development work in February, according to the company.

By August, that share had reached roughly one-quarter. Anthropic also says about 90% of its model research and development now involves collaboration with Claude. In that broader category, the model completes large portions of work under close human direction.

Those two numbers measure different levels of delegation. The 90% figure includes work where a person stays closely involved. The 26% figure covers assignments where Claude takes greater responsibility for execution after receiving the objective.

The distinction matters because neither number means Claude operates independently. The model does not choose Anthropic’s corporate goals, approve its own deployment, or control the complete training process. Human researchers still frame problems, allocate resources, inspect outcomes, and authorize consequential decisions.

Still, the reported shift is unusually fast. The initial disclosure says Claude moved from no leadership share to 26% within six months. That trajectory, rather than the current percentage alone, explains the attention.

Anthropic divides frontier-model development into engineering and research. Engineering includes writing software, maintaining infrastructure, and overseeing training systems. Research includes selecting experiments, interpreting results, and deciding which ideas deserve further testing.

Claude has become deeply involved in both categories. On engineering tasks, Anthropic says people increasingly provide goals without prescribing every implementation step. On research tasks, Claude can execute well-defined experiments and recommend subsequent actions.

A concrete incident illustrates the difference between ordinary code completion and delegated engineering. A routine upgrade had begun crashing tens of thousands of training jobs. An engineer supplied Claude with context and access to the affected cluster.

Claude tested environment settings, isolated an obscure debugging flag, reproduced the failure, and confirmed a fix. Anthropic says the investigation took about two hours. The company estimates that a human engineer would usually need two or three days.

This is not autonomous model creation. It is still significant because training infrastructure directly affects how quickly a laboratory can test and improve models. Compressing days of debugging into hours gives researchers more opportunities to run experiments and correct failures.

The company’s research account therefore describes a narrowing human role, not its disappearance. Claude increasingly handles implementation and experimentation. People retain authority over direction, evaluation, and release decisions.

That boundary defines the central tension. Anthropic has presented evidence of rapid automation inside AI development, while also documenting the human judgment that prevents the process from becoming self-directed.

The Coding Numbers Show Why Anthropic Is Sounding the Alarm

Claude’s contribution becomes more consequential when code generation, review, and experimentation begin reinforcing one another.

As of May 2026, Anthropic says Claude authored more than 80% of the code merged into its codebase. Before Claude Code entered research preview in February 2025, the share remained in the low single digits.

The company also reports a large increase in engineering output. During the second quarter of 2026, its typical engineer merged eight times as much code daily as in 2024. Anthropic attributes much of that increase to engineers directing and reviewing Claude rather than typing every change.

Lines of code are an imperfect productivity measure. More code can create maintenance costs, duplicate existing functions, or conceal poor design. Anthropic explicitly acknowledges that the eightfold increase almost certainly overstates the genuine productivity improvement.

However, output volume is only one piece of its case. Anthropic says intervention rates have declined across Claude Code sessions. Engineers reportedly correct, redirect, or take over from the model less often than they did one year earlier.

On Anthropic’s most open-ended engineering tasks, Claude’s reported success rate reached 76% in May 2026. That represented a 50-percentage-point increase over six months. A Claude-based evaluator judged whether sessions completed their assigned tasks without needing corrections.

That methodology deserves caution. An internal model judging another model does not provide the same assurance as an independent evaluation. Task composition can also change over time, complicating comparisons between monthly success rates.

Anthropic nevertheless reports that Claude-written code approached the quality of human-written code inside the company during 2026. Staff opinions were not unanimous. The company says many employees considered Claude’s code worse in late 2025 but roughly comparable by mid-2026.

Automated review adds another layer to the loop. Anthropic now uses Claude to inspect proposed code changes for defects and security weaknesses. A retrospective company analysis found that such reviews would have detected about one-third of bugs behind earlier production incidents.

This creates a development cycle with AI on both sides. One Claude session writes or modifies code. Another reviews the proposed change. Humans examine the result, decide whether it belongs in production, and investigate cases where automated checks disagree.

The pattern extends beyond programming. Anthropic tested models on an optimization task involving code that trains a small AI system. The goal was to improve speed without failing predetermined correctness checks.

Claude Opus 4 produced an average threefold speedup in May 2025, according to Anthropic. An internal model called Mythos Preview achieved roughly a 52-fold speedup by April 2026. The company says a skilled human researcher usually needed four to eight hours to reach a fourfold improvement.

Those figures show why Anthropic frames the trend as acceleration rather than simple automation. Faster implementation permits more experiments. More experiments produce additional evidence, which can guide the design of later models and tools.

The company’s own data remains the main source for these claims. No outside auditor has reproduced its internal task labels, success judgments, or productivity calculations. Readers should treat the reported trend as meaningful evidence, not as an independently established law of progress.

Even with that limitation, the direction is difficult to dismiss. Claude moved from suggesting snippets to editing repositories, operating tools, debugging live systems, and running experimental loops. Each step transfers another portion of AI development from manual execution to supervised delegation.

The Real Bottleneck Is Moving From Execution to Judgment

Claude is reducing the cost of doing research, but it has not eliminated the need to decide which research deserves to happen.

Anthropic’s strongest evidence concerns execution. Give Claude a measurable objective, access to relevant tools, and a way to test results, and the model can explore many possible solutions. This turns research implementation into a faster iterative process.

The harder problem is direction-setting. Researchers must choose questions that advance capabilities, improve safety, or resolve important uncertainties. They must also recognize misleading benchmarks, contaminated evaluations, and solutions that work only under narrow conditions.

Anthropic describes research taste as a continuing human advantage. This includes selecting valuable problems, deciding which evidence is credible, and abandoning approaches that have reached a dead end. Those choices shape the entire development process.

The company has tested whether Claude can recommend better next steps during realistic investigations. Researchers selected 129 moments when a human had taken an unproductive detour during earlier work. Different Claude models then proposed what should happen next.

A separate Claude model compared those recommendations with the eventual outcome. Anthropic’s Opus 4.5 reportedly beat the original human choice in 51% of cases during November 2025. Mythos Preview reached 64% in April 2026.

The result suggests improving tactical judgment, but it does not establish independent research leadership. Anthropic deliberately selected moments where the original human decision had room for improvement. A model with access to the completed session also served as evaluator.

The experiment therefore measures whether Claude can navigate selected decision points. It does not show that Claude can define a research program, recognize its broader consequences, or decide when a new model should be deployed.

A separate project tested broader autonomy in AI safety research. Multiple Claude agents investigated whether a weaker model could supervise a stronger one. They generated hypotheses, ran experiments, shared findings, and adjusted their approach.

Anthropic says two human researchers recovered about 23% of the gap between weak and strong supervision over roughly one week. The agents recovered 97% through 800 cumulative hours of work and substantial computing resources.

Yet the company identified important constraints. Humans chose the question and designed the scoring system. The result did not transfer cleanly to production-scale models. Those caveats separate an impressive experimental loop from autonomous scientific discovery.

The same pattern appears in everyday Claude Code use. Anthropic’s usage analysis found that users make about 70% of planning decisions in a typical session. Claude makes about 80% of execution decisions.

That division offers a clearer description than “AI builds itself.” People generally decide what should happen. Claude increasingly decides how to carry out the instruction, sometimes through long chains of file reads, code edits, tests, and command execution.

Domain expertise remains important because someone must recognize a plausible but incorrect result. An accountant can direct a reconciliation script by understanding accounting rules. An infrastructure engineer can identify whether a proposed fix creates unacceptable operational risk.

As execution becomes cheaper, judgment becomes more valuable. Researchers can supervise more experiments, but they also face more outputs requiring evaluation. The bottleneck moves rather than vanishes.

This shift affects how technical teams organize work. Writing code becomes a smaller part of the job. Specification, review, system design, verification, and accountability occupy a larger share.

Anthropic’s own hiring evaluations have encountered this transition. The company says performance engineers still handle difficult debugging and design decisions, even when Claude solves conventional coding tests. Its evaluation findings show that representative assessments become harder when AI can complete the mechanical portions.

The practical challenge is maintaining expertise when people perform less direct implementation. Reviewers need enough understanding to catch hidden errors. Yet constant delegation can reduce the hands-on experience that develops that understanding.

That is the unresolved human side of Anthropic Claude self-improvement. A faster model can expand what each researcher oversees. It can also make research systems harder for any single person to understand completely.

Anthropic’s Claim Pressures Every Frontier AI Lab

If Anthropic’s measurements are directionally correct, competing laboratories must match both the acceleration and the transparency surrounding it.

OpenAI, Google DeepMind, Meta, and other frontier developers use AI coding tools internally. Their exact methods and delegation levels differ, but all face the same competitive incentive. A laboratory that runs more useful experiments can improve models faster.

Claude’s reported contribution gives Anthropic a potential feedback advantage. Better models perform more engineering work. That work improves infrastructure and experimentation, which helps produce the next model.

The process does not need full autonomy to matter competitively. Human researchers can keep choosing goals while AI multiplies their execution capacity. Anthropic calls this compounding acceleration, even without a completely closed self-improvement loop.

That dynamic pressures competitors to deploy agents more aggressively. Waiting for perfect reliability carries an opportunity cost when another company can test more ideas, repair infrastructure faster, and reduce the time between model generations.

The resulting race creates a governance problem. Each laboratory controls the internal evidence showing how much model development has been automated. Outsiders cannot easily compare “collaboration,” “leadership,” task success, code quality, or researcher productivity across companies.

Anthropic has urged other developers to publish comparable measurements. It argues that common methods would help the public understand how close laboratories are to recursive self-improvement. Regular reporting could also reveal whether current growth curves are continuing or flattening.

The reported announcement arrived while Anthropic leaders were also calling for slower frontier development. That combination creates an unavoidable contradiction.

Anthropic is accelerating its own research with Claude while warning that acceleration might become difficult to control. Competitive pressure helps explain why the company does not simply stop using the technology.

Unilateral restraint could leave a cautious laboratory behind without slowing the overall field. Coordinated standards could reduce that disadvantage, but they require agreement across companies and governments with different economic and strategic interests.

The disclosure therefore serves several purposes. It informs the public, strengthens Anthropic’s safety argument, and advertises the capabilities of Claude. Those goals can coexist, but readers should recognize each one.

Competitors also have reasons to challenge Anthropic’s framing. A high share of AI-authored code does not measure model intelligence directly. Code volume says little about whether Claude originated a valuable research idea or merely implemented a detailed specification.

Different laboratories may categorize work differently. One company might label an agent-led debugging session as research leadership. Another might count the same task as an engineering tool used by a human researcher.

A shared measurement standard would need to separate planning, execution, verification, and authorization. It would also need independent auditing. Without those controls, companies could publish impressive percentages that remain impossible to compare.

The broader pressure extends beyond model laboratories. Enterprise engineering leaders will ask whether Anthropic’s internal workflow applies to them. Many will pursue similar systems that connect coding agents with repositories, tests, deployment tools, and operational data.

The answer depends on verification quality. Anthropic builds models and employs specialists capable of reviewing unusual failures. A company without comparable expertise can automate code production faster than it improves its ability to evaluate that code.

AI-generated output can therefore widen differences between organizations. Teams with strong tests, documentation, observability, and review practices can delegate more safely. Teams with weak foundations risk producing errors at higher speed.

This is also why searchable institutional context matters. An agent needs access to requirements, prior decisions, and technical constraints. A maintained AI knowledge base can help people inspect the reasoning around delegated work, although it cannot replace technical validation.

For enterprise buyers, Anthropic’s announcement is not a command to automate everything. It is evidence that AI-assisted development is moving from individual suggestions toward supervised workflows. The competitive question concerns how much responsibility an organization can verify, not merely how much it can delegate.

What Anthropic’s Numbers Still Do Not Prove

Anthropic has shown rapidly expanding assistance, but it has not shown a model independently designing, training, and approving its successor.

The phrase “building itself” compresses several distinct activities. Writing application code differs from selecting a model architecture. Running an experiment differs from deciding whether its result supports a major training investment.

Claude performs meaningful work across that spectrum, according to Anthropic. However, people still provide goals, evaluation criteria, infrastructure access, and release authority. Those controls prevent the current process from qualifying as fully recursive self-improvement.

Measurement quality is another concern. Most central figures come from Anthropic’s internal systems. The company created the task categories, collected the sessions, and evaluated many outcomes with Claude-based judges.

Model judges can process large datasets consistently, but they can inherit blind spots from the systems they evaluate. A session may appear successful because tests passed while introducing maintenance, security, or architectural problems outside the test scope.

Code authorship is similarly ambiguous. Claude might generate an entire file after extensive human planning. Alternatively, a researcher might supply a short objective while the model determines the implementation. Both files count as AI-authored, although they represent different levels of independence.

The eightfold output figure also measures merged code, not verified economic value. An engineer who generates more code may resolve more problems. The same engineer might also create more review work and larger systems that become expensive to maintain.

Independent research offers useful context without resolving Anthropic’s internal claims. One 2026 working paper examined 7.8 million Claude-co-authored commits and a panel of 5,838 developers. The study associated adoption with work across more repositories and technologies.

Its developer study suggests that coding agents can expand the range of tasks people attempt. However, public GitHub activity does not measure Anthropic’s internal model research or prove recursive improvement.

Reliability also varies by task. Structured problems with clear tests favor automated iteration. Open-ended research questions often lack an objective score, allowing plausible but weak conclusions to survive longer.

Compute and physical infrastructure impose further limits. Faster agents can propose and run more experiments, but training frontier models requires chips, energy, networking capacity, and carefully managed data. Intelligence is not the only scarce resource.

Human review can become a bottleneck as output grows. If Claude generates changes faster than specialists can examine them, Anthropic must either slow deployment or rely more heavily on automated evaluation.

That creates a circular verification problem. Claude writes code, another Claude reviews it, and Claude-based systems judge whether tasks succeeded. Human specialists remain the final backstop, but their visibility can decline as the automated chain grows longer.

Security raises the stakes. An agent with repository, cluster, and command access can solve difficult operational problems. The same access increases the consequences of faulty reasoning, compromised context, or manipulated instructions.

Organizations adopting similar workflows need narrow permissions, complete logs, reproducible tests, and clear escalation paths. They also need people who remain accountable for production changes, regardless of who generated the code.

Anthropic itself does not claim that Claude has crossed the autonomy threshold. Its more defensible argument is that the distance is shrinking. Engineering execution has become highly automated, while research judgment shows early but limited improvement.

That claim is important enough without exaggeration. A system does not need to become an independent scientist before it changes the speed and economics of AI development.

Three Signals Will Show Whether the Loop Is Really Closing

The next stage depends on measurable changes in research leadership, independent validation, and human control over consequential decisions.

The first signal is Claude’s share of model research and development leadership. Anthropic reported a rise from zero in February to 26% in August 2026. Future disclosures should show whether that curve continues, plateaus, or reverses.

A continuing increase would strengthen the compounding-acceleration case, especially if tasks become less specified. A plateau would suggest that implementation improved faster than research judgment. A decline could indicate measurement changes, harder assignments, or reliability problems.

The definition of leadership must remain stable. Otherwise, a higher percentage could reflect recategorized work rather than greater independence. Anthropic should publish examples across engineering, experimentation, interpretation, and strategic planning.

The second signal is independent replication. Outside evaluators need access to representative tasks, scoring criteria, and results without receiving sensitive model-development secrets. Comparable tests across laboratories would be more informative than isolated company percentages.

Replication should examine code quality over time, not only task completion. It should measure defects, security issues, maintenance burden, and the frequency of human intervention. Research tasks should test whether agents select productive directions before seeing eventual outcomes.

Independent evidence would either reinforce or weaken Anthropic’s account. Similar results across different evaluators and environments would support the claim of a broad capability shift. Large gaps would show that internal workflows or evaluation choices drive much of the reported performance.

The third signal is how Anthropic handles human oversight as output grows. The company already identifies review as a potential bottleneck. Its response will reveal whether people retain meaningful control or increasingly approve decisions through automated summaries.

Watch for changes in review requirements, access controls, deployment procedures, and incident reporting. Also watch whether Claude begins choosing research goals rather than only implementing them.

A genuine closing of the loop would require more than a higher code-authorship percentage. Claude would need to propose valuable research directions, design credible evaluations, interpret ambiguous evidence, and improve its successor with limited human guidance.

Even then, authorization remains separate from capability. A model might become technically able to direct more research while Anthropic deliberately preserves human approval. Governance choices will determine whether potential autonomy becomes operational autonomy.

For developers, the immediate lesson is practical. Coding is shifting toward specification, orchestration, testing, and review. Engineers who understand systems deeply can delegate more while recognizing when an agent’s confident output is wrong.

Enterprise buyers should ask what vendors measure when they claim autonomous work. Completion rates alone are insufficient. Buyers need evidence about supervision, reversals, defects, auditability, and responsibility after failures.

Knowledge workers face a similar transition. AI can perform more execution, but people still need reliable context and a record of important decisions. Tools for knowledge blending become useful when they help users trace evidence instead of merely generating another answer.

Anthropic Claude self-improvement is therefore real in a limited, consequential sense. Claude already accelerates the engineering and experimental work that produces later Claude models. The company has not shown an autonomous system independently building its successor.

The gap between those statements is where the next chapter will unfold. Track the leadership percentage, demand independent validation, and examine who still makes irreversible decisions. If those three signals move together, “AI building AI” will describe an operating system rather than a headline.

The question for every organization is no longer whether AI can produce more work. It is whether people can verify that work, preserve the expertise needed to challenge it, and remain accountable as the automated chain expands.