On September 17, 2026, Anthropic published Measurements for understanding the pace of AI development inside frontier labs: three public measurements — (1) how much of AI R&D is performed by AI itself; (2) how well actions of AI agents are overseen; (3) how compute is allocated — plus a snapshot from inside Anthropic. Numbers would be expected to shift if there were coordination on pacing the frontier, as called for by CEO Dario Amodei (reference only; Amodei essay HOLD / do not re-ship). Plans to embed independent third-party evaluators with access comparable to internal risk teams to verify safety practices, report incidents, and monitor metrics (planned, not claimed complete). Complements capability evaluations / RSP risk reports; references Advanced AI Framework (AAIF) policy proposal. (1) Anthropic R&D Automation Index using Epoch AI Automation Level AL0–AL5. As of August 2026 (Anthropic self-report): Claude not fully autonomous (AL5) for any measured subset; Claude “leads” (AL4) 26% of Anthropic’s AI R&D work; share at or above “AI collaborates” (AL3+) above 90%. (2) Agent oversight (most-used internal platform; Aug 2026): ~30,000 agents doing research/engineering at any one time; online monitors 100% coverage before execution; offline 100% after; analyzed >1 billion decisions Aug 2026: 0.002% blocked (~1 in 47,000); offline flags ~1–2 per thousand; ~100k transcripts/week flagged then classifiers; ~50/week highest priority to human review. (3) Compute allocation (July 13–20 snapshot, verified in pack dump): about 6% of compute that went to AI R&D allocated toward safety; about 12% of compute that went to AI-driven AI R&D allocated toward safety; deliberately conservative; does not account for safeguards classifiers. FLAG: all internal metrics = Anthropic self-report; third-party verification planned not yet claimed complete. This is a NEW primary on measurement tools — not a re-ship of We Must Pace the Frontier. Primary verified via signal/sources/EU-SIGNAL-20260918-001/evidence-anthropic-pace-measurements.md.
Anthropic is publishing process-side transparency metrics — R&D automation, agent oversight, and compute-to-safety share — that sit beside capability evals and RSP reports. Track as measurement tools + a self-reported internal snapshot, with third-party verification still ahead. Attribute all figures as Anthropic self-report; do not conflate with or re-ship the Amodei pacing essay.
- 01
Sep 17, 2026: Three proposed public measurements — AI-led R&D share, agent oversight, compute allocation — plus Anthropic internal snapshot.
- 02
Automation Index (Aug 2026): No AL5 subset; Claude “leads” (AL4) 26% of Anthropic AI R&D; AL3+ >90% (Anthropic).
- 03
Agent oversight (Aug 2026): ~30k agents on top internal platform; 100% online + offline coverage; 0.002% blocked (~1/47k); offline ~1–2/1k flags; ~100k/week → classifiers; ~50/week to humans.
- 04
Compute (Jul 13–20): ~6% of AI R&D compute → safety; ~12% of AI-driven AI R&D compute → safety; conservative; classifiers excluded from these %.
- 05
Third parties: Independent evaluators with internal-comparable access planned — not claimed complete.
- 06
Context: Complements RSP capability/risk reporting; references AAIF policy proposal; Amodei pacing call = attribution only.
- 07
FLAG: All internal metrics = Anthropic self-report; not a re-ship of “We Must Pace the Frontier.”
Last verified: 2026-09-18T13:40:00Z · Speaker predictions remain attributed to the speaker. ElevateU analysis is labeled separately from source-supported fact.