Role here: evidence artifact for pace / automation claims
Metric 1 - AI-led AI R&D
Item
Value
Instrument
Anthropic R&D Automation Index
Scale
Epoch AI Automation Levels AL0-AL5
As of
August 2026
AL4 ("leads") share
26% of Anthropic AI R&D work
≥ AL3 ("collaborates")
>90%
AL5 fully autonomous
none in measured subsets
Appendix method: sample staff weeks in July 2026 → ~15,000 tasks → frozen tree (542 nodes / 378 leaves) → Claude research + Claude judge ratings → person-time weights. Human agreement check reported in-appendix.
Metric 2 - Oversight of AI agents
Item
Online monitors
Offline monitors
Role
Block/steer before execution
Review after the fact
Coverage (stated)
100% of actions on the measured platform
100% ingested after the fact
Platform scale
~30,000 agents on most-used internal research/engineering platform
same
August signal
>1B decisions; ~0.002% blocked (~1 in 47,000)
~1-2 / 1000 transcripts flagged; ~50/week high-priority human escalations
Metric 3 - Compute allocation
Item
Value
Window
13-20 July (one week snapshot)
Safety share of AI R&D compute
~6%
Safety share of AI-driven AI R&D compute
~12%
Coding rule
Dual-purpose tokens counted as non-safety (conservative)
Note
Safeguards-classifier compute excluded from these shares
Appendix pointers
AL3 vs AL4 concrete pipeline-fix vignette in footnotes
Compute classifier prompt excerpt for safety vs R&D labeling
Plan to embed third-party evaluators with internal-comparable access
Reading rule
Cite numbers with as-of dates and definitions. The 26% figure is AL4 on a person-time-weighted AI R&D basket. It is not headcount, revenue, or total company work.
The Frontier, “Anthropic's three pace metrics and appendix”, 20 Sept 2026
TLDR
Automation Index, agent oversight table, and safety compute shares from the Institute post, with method notes.
Plain text
Frontier Survey. “Anthropic's three pace metrics and appendix.” The Frontier. 20 Sept 2026. https://www.thefrontier.dev/articles/anthropic-pace-metrics
BibTeX
@misc{frontier_anthropic_pace_metrics_2026,
title = {Anthropic's three pace metrics and appendix},
author = {{Frontier Survey}},
howpublished = {The Frontier},
year = {2026},
month = sep,
url = {https://www.thefrontier.dev/articles/anthropic-pace-metrics}
}
Full text may be reprinted with canonical link and byline.
Frontier Red Team's unreleased research model, under human prompting, produced five disclosed chains and helped drive 62 Chrome CVEs into the 156 release.
Explore related tags
Jump to other Frontier pieces on the same subjects.
