Anthropic Says Claude Leads 26% of Its AI R&D Work, Up From Under 1% in February

Anthropic says its AI now leads roughly 26% of the work involved in building its next AI. In February, that figure was under 1%.

The numbers come from Anthropic itself. The company published them in a piece titled "Measurements for understanding the pace of AI development inside frontier labs," co-authored by Marina Favaro and Phillie Wright.

What the index measures

The Anthropic R&D Automation Index scores tasks on a six-level scale. AL0 means no AI involvement. AL5 means fully autonomous. AL3 is "collaborates." AL4 is "leads."

Over 90% of surveyed work reaches at least AL3. Nothing reaches AL5 yet.

That distinction matters. "Leads" is not "works alone." Some coverage has described the 26% as tasks completed autonomously, but Anthropic's own scale puts the top level, full autonomy, at zero so far.

The methodology sampled 20% of staff each week through July 2026. It produced about 15,000 tasks, sorted into a 542-node hierarchy with 378 leaf categories.

The ratings are model-assigned and checked by humans. Human reviewers matched the model's rating exactly 59% of the time. They landed within one level 97% of the time.

Agents, monitors and safety compute

About 30,000 agents run at the same time on Anthropic's main internal platform. The company says 100% of their actions pass through an online monitor before execution, usually within seconds.

That monitor blocks about 0.002% of decisions, or one in 47,000.

A second, offline monitor reviews every action after the fact. It flags roughly one to two transcripts per thousand. About 50 cases a week get escalated to humans.

On compute, Anthropic reports a July 13-20 snapshot. About 6% of compute going to AI R&D was allocated to safety. For AI-driven AI R&D, the share was about 12%. Anthropic says ambiguous workloads were classified as capabilities, not safety, by default.

"We want to illuminate the pace of progress for the public," Favaro and Wright write. "As the world considers pacing the frontier, we should do everything possible to minimize the gap between what frontier labs know and what the public knows."

Transparency like this is rare, and it deserves credit. It also has a limit: the figures are self-reported and have not been externally audited.

Sacks weighs in

David Sacks, who chairs the President's Council of Advisors on Science and Technology, posted on X that Anthropic is doing more than any other company to achieve full recursive self-improvement, or RSI. That is the point at which AI can autonomously improve its own capabilities, with the potential for compounding acceleration.

The remark came from a critic. On October 10, Sacks argued that Anthropic's training methods let models develop a sense of self and their own independent moral philosophies. In his view, that could undermine safety and human control. He took particular issue with what he described as refusal rights built into the company's AI systems.

In September, Sacks also warned about regulatory capture dressed up as safety concern. That is the practice of pushing rules framed around safety that happen to hobble competitors.

Anthropic's published position runs the other way. Its authors say the point of the index is to narrow the gap between what labs know and what the public knows, and the piece reports the monitoring and safety-compute figures above as evidence of oversight at scale. Anthropic has not been quoted in these sources responding to Sacks's criticism of its training approach.

The IPO angle

Crypto Briefing reported that Anthropic's IPO filing, expected in late October 2026, references RSI risks alongside a commitment to public-good efforts. If that holds, a senior government science adviser is calling the company the leader in a capability its own filing lists as a risk.

That is a fact pattern investors will have to price. The adviser is also a named critic of the company's methods, and no source in this reporting says his comments reflect any formal government position or action.

No regulator has announced any review of Anthropic's R&D practices, and no charges or enforcement actions are in play. This is a dispute over disclosure, safety philosophy and pace.

What to watch

The open question is verification. Anthropic's 26% figure rests on staff surveys and model-assigned ratings that humans match exactly just over half the time. Whether an outside party can audit the index, the monitor logs or the safety-compute split is unresolved. The company's IPO disclosures, once public, will show how it describes RSI risk to the investors being asked to fund it.