Quick Answer
Anthropic says Claude led 26% of measured work building successor AI models in August 2026, up from under 1% in February. The company also says no measured work is fully autonomous, while an oversight monitor reviewed agent actions before execution. These are company-reported internal measurements, so readers should treat them as a transparency signal, not independent proof of capability.
Key Takeaways
- Anthropic says Claude led 26% of its measured AI research and development work in August 2026.
- Claude’s share of work at the “leads” level rose from under 1% in February to 26% in August.
- More than 90% of measured work reached at least Anthropic’s “AI collaborates” level under close human direction.
- Anthropic says its monitor reviewed 100% of actions from roughly 30,000 internal agents before execution.
- The company says it plans to publish these measurements regularly so outside observers can track progress.
What did Anthropic disclose about Claude building next AI models?
Anthropic says Claude led 26% of the company’s measured AI research and development work in August 2026. The company defines that “leads” category as work where Claude can complete most of a task from a high-level prompt while a human supervises the process, according to Anthropic’s methodology disclosure.
The distinction matters because “leads” does not mean Claude independently chooses goals, deploys changes, or operates without review. Anthropic says Claude was not fully autonomous for any measured subset of AI R&D work. The company labels full autonomy as AL5 on its automation scale, and none of the measured work reached that level.
Anthropic’s disclosure focuses on internal development activity rather than a new consumer Claude feature. The separate Claude workplace features launch concerns user-facing tools, while the September 17 announcement describes how Anthropic measures AI assistance inside its own research and engineering processes.
How does Anthropic define Claude’s “leads” level of AI work?
Anthropic defines Claude’s “leads” level as a supervised workflow where the model completes most of a task after receiving a high-level human prompt. A person remains responsible for directing the task and supervising the work, which separates the category from fully autonomous operation.
Anthropic also says more than 90% of its measured AI R&D work reached at least the “AI collaborates” level in August 2026. That lower threshold means Claude can complete large portions of work under close human direction, rather than simply offering isolated suggestions or short answers.
The two categories describe different degrees of model contribution. “AI collaborates” indicates substantial assistance under close direction, while “leads” indicates that Claude carries more of the task after a broad instruction. The practical reading is that Anthropic is measuring workflow delegation, not claiming that Claude has replaced the researchers and engineers who set objectives, review outputs, and approve changes.
| Anthropic measurement level | What Anthropic says Claude can do | Human role |
|---|---|---|
| AI collaborates | Complete large portions of work | Provide close direction and supervision |
| Leads | Complete most of a task from a high-level prompt | Supervise the work and retain responsibility |
| AL5, fully autonomous | Operate without the supervision described above | No comparable supervised role |
How quickly did Claude’s role in AI research change?
Claude’s share of measured work at Anthropic’s “leads” level rose from under 1% in February 2026 to 26% in August 2026. The change occurred across a six-month period, based on Anthropic’s internal measurement framework and the company’s reported assessment of its own research and engineering work.
The increase matters because it suggests that Anthropic researchers are assigning Claude larger portions of some development tasks than they did earlier in the year. The figures do not show that every research task can be delegated, and they do not establish how Claude performs on work outside Anthropic’s systems or evaluation process.
Anthropic’s updated chart in its recursive self-improvement disclosure adds task-success figures. Anthropic says Claude Code session-success rates across four task categories converged at roughly 88% to 92% by September 2026, and that the success rate for open-ended tasks rose from about 26% to about 91%. These are self-reported measurements of Anthropic’s internal task categories rather than an independent industry benchmark, and a successful session does not establish that every output is correct, secure, or suitable for deployment.
Anthropic also cited an outside trend showing that the duration of tasks AI agents can reliably complete has doubled roughly every 4 months, compared with an earlier trend of roughly 7 months. The nonprofit research group METR defines this task-completion time horizon as the human-expert task duration at which a model is predicted to succeed at a specified reliability level. METR’s time-horizons methodology focuses primarily on software engineering, machine-learning, and cybersecurity tasks, which limits how broadly readers should apply the trend.
Outside readers should also separate a larger share of supervised work from a direct measure of scientific capability. Anthropic’s figures describe the proportion of work performed at a defined level of automation. They do not provide a universal benchmark score, an independent audit of output quality, or a guarantee that the same workflow would work at another AI company.
How does Anthropic oversee the agents doing research and engineering work?
Anthropic says roughly 30,000 agents were conducting research and engineering work at any one time on its most-used internal platform in August 2026. The company says an online monitor reviewed 100% of those agents’ actions before execution, placing an automated control layer between proposed actions and the company’s internal systems.
Anthropic says the monitor blocked 0.002% of more than 1 billion decisions analyzed, which is about 1 in 47,000 decisions. A low block rate can mean the system permitted most actions, but it does not establish that every permitted action was correct, safe, or useful. Monitoring systems depend on the rules, signals, and failure modes that their designers choose to detect.
The oversight model is relevant beyond Anthropic because AI agents can take actions rather than merely generate text. Readers following questions about model safeguards may recognize similar concerns in AI misalignment reporting, where the central issue is whether developers can observe and control behavior before it creates a harmful result. Anthropic itself says that AI systems able to fully build their successors would make securing, monitoring, and shaping AI behavior much more important.
Why did Anthropic disclose its AI agent monitoring figures?
Anthropic introduced three proposed public measures for tracking frontier AI development: AI-led R&D, agent oversight, and compute allocation. The company presented the measures on September 17, 2026, as a way to make progress in AI development more visible to outside observers.
Anthropic says it intends to publish the measurements regularly so external observers can track how quickly AI contributes to building successor systems. Reuters reported that Anthropic plans recurring publication of the measurements, which could make changes in the company’s internal automation levels easier to compare over time.
Regular disclosure would be more useful than a single snapshot because the underlying numbers can change quickly as models, tools, and internal policies change. That said, the measurements remain company-produced figures. The practical response is to watch whether Anthropic publishes consistent definitions, historical comparisons, and enough methodological detail for outsiders to understand what each percentage measures.
What does Anthropic’s compute allocation figure show?
Anthropic says about 6% of the compute used for AI R&D went to safety work during a one-week snapshot from July 13 through July 20, 2026. The company describes that estimate as deliberately conservative, meaning the figure may exclude some work that could reasonably be considered related to safety.
The compute figure matters because advanced AI development requires substantial computing resources, and resource allocation can show what a company is prioritizing operationally. A percentage of compute is not a complete measure of safety investment, however. Safety work can also involve personnel, policy development, evaluations, monitoring systems, and processes that do not map cleanly to a single compute category.
Anthropic’s proposed metric therefore offers a narrow but concrete view of internal priorities. Readers should avoid interpreting 6% as a complete safety score or a direct comparison with another company, especially when other organizations may classify research, evaluation, and security work differently.
What can readers conclude from Anthropic’s Claude automation data?
Anthropic’s data indicates that Claude has become more involved in supervised work used to build future AI systems, but the disclosure does not show full autonomy. The company says Claude led 26% of measured AI R&D work in August and remained below its fully autonomous AL5 category across every measured subset.
The most important limitation is that Anthropic designed the measurement system and reported the results itself. The Associated Press also covered the company’s disclosure in its report on Anthropic’s internal AI development measurements, but outside coverage does not turn company-reported metrics into independently validated benchmarks.
For most readers, the practical conclusion is that AI development is moving toward more agent-assisted workflows, with human supervision and automated monitoring still presented as core controls. Comparisons with other systems should remain cautious. Related developments, such as background reasoning for voice tasks, may show broader industry interest in more capable AI workflows, but they do not measure the same internal research process.
FAQ
Does Claude build Anthropic’s next AI models by itself?
Anthropic says Claude does not build its next AI models by itself because no measured AI R&D work reached the company’s fully autonomous AL5 level. Claude can lead most of a task from a high-level prompt while a human supervises the process.
What does Claude leading 26% of AI R&D work mean?
Claude leading 26% of AI R&D work means Anthropic measured Claude completing most of those tasks under human supervision in August 2026. The figure does not mean that 26% of Anthropic’s research operates without people.
How many AI agents does Anthropic use internally?
Anthropic says roughly 30,000 agents were performing research and engineering work at any one time on its most-used internal platform in August 2026. The company says an online monitor reviewed each agent action before execution.
Did Anthropic’s oversight monitor block many agent actions?
Anthropic says its oversight monitor blocked 0.002% of more than 1 billion decisions analyzed, or about 1 in 47,000 decisions. The low reported rate does not prove that every approved decision was correct because monitoring quality depends on what the system can detect.
Will Anthropic publish Claude automation measurements again?
Anthropic says it intends to publish the measurements regularly so outside observers can follow how quickly AI contributes to successor-system development. Future reports will be most useful if Anthropic keeps the methodology consistent and explains any changes in its definitions.