# Anthropic Says Claude Now Leads 26 Percent of Its Own AI Research Work

The company published three internal metrics, including a compute snapshot showing roughly 6 percent of its research and development compute went to safety work.

- Published: 2026-09-19T06:16:00.678Z
- Canonical: https://polylog.news/ai/2026-09-19/anthropic-says-claude-now-leads-26-percent-of-its-own-ai-res
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Anthropic](https://www.anthropic.com/institute/measuring-pace-of-ai-development), [CNBC](https://www.cnbc.com/2026/09/17/anthropic-shares-3-metrics-to-help-ai-companies-monitor-development.html), [Polylog editors](https://polylog.news)

Anthropic published [three measurements](https://www.anthropic.com/institute/measuring-pace-of-ai-development) of what is happening inside its own building: how much of its research and development (R&D) Claude performs, how many autonomous agents run in its systems, and how its compute is split. The headline number is an internal index in which Claude "leads" about 26 percent of model R&D tasks as of August, meaning the model completes a task end to end from a high-level prompt while a human reviews the result. [CNBC reported](https://www.cnbc.com/2026/09/17/anthropic-shares-3-metrics-to-help-ai-companies-monitor-development.html) the same figures and noted Anthropic's caveat that no subset of the work runs fully autonomously.

The trajectory is the part worth arguing about. The same index sat [below 1 percent earlier in the year](https://www.implicator.ai/anthropic-claude-leads-26-percent-ai-research/), and roughly 30,000 agents now run research and engineering tasks simultaneously on the company's main internal agent platform, a figure also circulated by [AI Post](https://t.me/aipost/8187). Anthropic says its monitoring system reviewed every action those agents took before execution and blocked about one in 47,000 of more than a billion decisions.

Two readings are available and the data does not settle between them. One is that a self-reported index defined by the company that benefits from it will drift upward as task definitions loosen. The other is that recursive automation of research is genuinely compounding, in which case the relevant constraint on lab progress becomes compute and review capacity rather than headcount. The compute snapshot, taken between July 13 and July 20, showed roughly 6 percent of R&D compute allocated to safety.

Anthropic released its methodology and invited other labs to publish comparable numbers. None has yet. Until a second lab reports on the same definitions, the index is a vendor disclosure, not an industry measurement, and its main value is that it is falsifiable at all.

## What this means

If model-led research keeps compounding, the binding input to frontier progress shifts from research staff to compute and to the review capacity that checks agent output, which favors labs with secured accelerator supply and penalizes those whose advantage was hiring. The disclosure also gives regulators a template: a metric defined by a lab, published voluntarily, that can be written into reporting requirements later. Rivals that decline to publish comparable figures will be asked why.

## What to watch

- Whether OpenAI, Google DeepMind, or a Chinese lab publishes an R&D automation figure on Anthropic's definitions, which is the only way to tell an industry trend from one company's accounting.
- Whether the share of research compute going to safety moves in later snapshots, since a falling share alongside a rising automation index would be the clearest sign that oversight is being outrun.
- Whether Anthropic's headcount growth in research roles slows, which would indicate the automation is substituting for hiring rather than adding to it.