Anthropic: Claude now leads 26% of its AI R&D

Anthropic published its R&D Automation Index and related pace metrics. As of August 2026, Claude “leads” 26% of Anthropic’s AI research and development work on Epoch AI’s automation scale — meaning it can finish most of a task from a high-level prompt while a human supervises — up from about 1% in March. AI collaborates on more than 90% of R&D; Claude is not fully autonomous in any measured area. The post also reports about 30,000 agents on its main internal platform, near-complete action monitoring, and a conservative about 6% safety share of AI-R&D compute in a July sample week.

SAFETY desk — the clearest same-day public window into how fast a frontier lab’s own model is taking over the work of building the next model, with oversight and safety-compute numbers attached.

As of August 2026, Anthropic says Claude “leads” 26% of the company’s AI R&D work. “Leads” here is Automation Level 4, or AL4, on a scale from Epoch AI, an independent nonprofit that tracks AI progress: the model can finish most of a task end-to-end from a high-level prompt while a human supervises. The share of work at or above “AI collaborates” — AL3, large chunks of work under close human direction — is above 90%. Claude is not fully autonomous, AL5, for any measured subset. File those 26% / 90%+ / no-AL5 lines as Anthropic’s. This desk did not rate a task.

The scale, as Anthropic uses Epoch AI’s wording, runs from AL0 (no AI involvement) to AL5 (AI operates fully autonomously, with no human in the loop). AL4 is not unsupervised takeover: a person still reviews and decides whether the work ships. File that AL0-to-AL5 picture as Anthropic / Epoch. This desk is not inventing a cross-lab ranking or a claim that other labs published the same index.

Contribution rose rapidly from about 1% “leads” in March 2026, Anthropic says. Reuters’ same-day report notes the same March-to-August jump. File that ~1% starting point as Anthropic’s, with Reuters as same-day confirm. This desk did not reconstruct the monthly series beyond what the company printed.

About 30,000 agents were doing research and engineering work on Anthropic’s most-used internal platform at any one time in August, the company says. Those measurements cover that platform only. Anthropic says 100% of those agents’ actions pass an online monitor before they run and an offline monitor after; over a billion decisions that month were analyzed, and about 0.002% — about 1 in 47,000 — were blocked by the online monitor. File the 30,000 / 100% / billion / 1-in-47,000 lines as Anthropic’s. This desk did not watch a monitor queue.

Compute snapshot, still company: for the week of July 13–20, about 6% of compute going to AI R&D was safety work, and about 12% of compute going to AI-driven AI R&D was safety. Anthropic calls those figures deliberately conservative — work that advanced capabilities as much as safety was counted as capability, and safeguard classifiers were left out. File the 6% / 12% / one-week / conservative lines as Anthropic’s. This desk did not meter a cluster.

Anthropic says it plans to embed independent third-party evaluators from several organizations, with access comparable to internal risk-assessment teams, to verify safety practices, report incidents, and monitor metrics such as these. File that third-party-evaluator plan as Anthropic’s. This desk is not inventing a signed inspector contract, a start date, or named evaluator firms. None were printed as a closed roster.

Plain English for the rest of the card: R&D = research and development, here the work of building the next AI model. R&D Automation Index = Anthropic’s prototype score for how much of that work Claude already does. Epoch AI = an independent nonprofit that tracks AI progress; Anthropic used its Automation Level scale. AL3 / collaborates = the model does large chunks of work under close human direction. AL4 / leads = the model finishes most of a task from a high-level prompt while a human supervises. AL5 / fully autonomous = no human in the loop — a level Anthropic says it has not reached in any measured subset. online monitor = a real-time check that can block an agent action before it happens. offline monitor = an after-the-fact review of what agents already did. “Leads” is not unsupervised takeover, and this is not a claim that recursive self-improvement — a model fully building its successor alone — has arrived.

PRIMARY here: Anthropic Institute’s 17 Sep 2026 post “Measurements for understanding the pace of AI development inside frontier labs” — Tier A PRIMARY company source, the original record. Reuters is same-day independent Tier B confirm of the 26% “leads” figure, the >90% collaborates line, the no-AL5 read, the March ~1% jump, the ~30,000 agents, the 1-in-47,000 block rate, and the 6% / 12% safety-compute snapshot. Bloomberg is same-day Tier B corroboration of the 26% research-and-development line — not a substitute primary, and not a source for invented AL5 claims, cross-lab comparisons, or IPO timing. The three measurements, the AL0–AL5 / Epoch scale, the August 26% AL4 / >90% AL3 / no-AL5 snapshot, the March ~1% starting point, the ~30,000 agents and 100% monitor coverage, the billion decisions / 0.002% / 1-in-47,000 block rate, the July 13–20 6% / 12% conservative compute shares, and the third-party-evaluator plan are Anthropic-attributed. NOT claimed: that recursive self-improvement is here, that Claude runs the company, that AL5 was reached, a cross-lab ranking, an IPO date, that this desk reran the index or sat on a monitor, a stock tip, or investment advice. Distinct from the already-filed openai-misalignment-reporting-framework, king-charles-ai-summit-dumfries, novo-anthropic-claude-drug-discovery, claude-cowork-chat-merge, amodei-pace-the-frontier, anthropic-threat-intelligence-september-2026, anthropic-september-2026-threat-report, and anthropic-intelligence-targeting-conventional-weapons.