Measurements for understanding the pace of AI development inside frontier labs

An Anthropic Institute post (September 2026) proposing three measures that frontier developers could publish, with a first snapshot from inside Anthropic, all produced with Claude models and not externally verified; an R&D Automation Index judges that Claude "leads" 26% of Anthropic's AI R&D work as of August 2026 and operates fully autonomously on none, internal agent monitoring blocked about 1 in 47,000 agent decisions in August, and about 6% of AI R&D compute in one July week went to safety work.

Marina Favaro and Phillie Wright (The Anthropic Institute) · 2026-09-17 · original

As AI systems take on more of the work of building their successors, people outside the labs have little visibility into how fast that is happening. This post from the Anthropic Institute, published on 17 September 2026, proposes three measures of how models are built, as distinct from capability evaluations of what they can do: how much AI research and development (R&D) is done by AI, how well AI agents working inside a lab are overseen, and what share of compute goes to safety. For each it describes a prototype, reports a snapshot from Anthropic, and says what any developer "could report today". The post places this in the context of calls to slow frontier development, citing an essay by Anthropic's CEO, and says the measures would give "the public, third parties, and governments better visibility". It names Marina Favaro and Phillie Wright as co-authors, with research direction from Jack Clark.

The web page is the only document. Methods are in its Appendix; numbers are located below by the post's section headings. The Automation Index chart prints its monthly values as labels. The post says it published the oversight measurements in Anthropic's August 2026 Risk Report; see "Comparison with the August 2026 Risk Report" below.

Measure 1: how much AI R&D Claude does

Method

The Anthropic R&D Automation Index rates how automated each kind of AI R&D work at Anthropic is, using a six-level automation scale proposed by Epoch AI: AL0 (no AI involvement), AL1 (minimal), AL2 (AI assists), AL3 (AI "collaborates", doing large chunks of work under close human direction), AL4 (AI "leads", completing most of a task end to end from a high-level prompt while a human supervises) and AL5 (fully autonomous, no human in the loop) ("(1) Measuring AI-led AI R&D"). Footnote 1 illustrates the levels with fixing a broken nightly data pipeline: at AL4, Claude finds and fixes the fault and writes it up but does not deploy; the engineer decides whether it ships.

The Appendix ("Measuring AI-led R&D") describes three steps, each done by Claude models:

  1. The task list. For each week of July 2026, 20% of staff in each department in the model R&D loop were sampled at random. A Claude research agent reviewed each sampled person's week from Slack and internal documentation and listed their tasks, giving about 15,000 granular tasks. Claude organized them into a tree of 542 nodes, 378 of them leaves (examples: "eval platform defect diagnosis and fixes", "serving incident postmortems"). The tree is frozen so that later measurements use the same basket of work.
  2. The ratings. For each node, a Claude agent researched how the work is done across the company, and "An independent Claude judge" read that evidence and assigned an automation level. For a given month, the rating agents see only evidence from that month or earlier.
  3. The weights. Each sampled person contributes one unit of weight per week, split evenly across the tasks they worked on, as a proxy for importance. The post calls this "a crude approximation".

Results

As of August 2026, Claude was not operating fully autonomously (AL5) on any measured subset of AI R&D work, "leads" (AL4) 26% of it, and the share at AL3 or above was "above 90%" ("What we found"). The chart (Claude now leads 26% of model R&D work, Index v2026.07, with 90% measurement intervals) prints the AL4 share by month in 2026: 1% in March, 3% in April, 12% in May, 14% in June, 22% in July and 26% in August, with almost none before March. The chart's series starts in August 2025; read from the chart, the share at AL3 or above rose from about 4% then to about 96% in August 2026. Because the task basket is frozen from July 2026 and ratings for each month use evidence up to that month, the earlier months are, on this note's reading, retrospective ratings of the July 2026 basket made in 2026. The post does not describe how the 90% intervals were computed. In the first archived copy of the page (Wayback Machine, 17 September 2026, 20:58 UTC), the chart's alternative text ended "If the trend continues, that share could reach 80% by the end of 2026."; by the capture of 29 September 2026 that sentence had been removed, and the live text ends at "up from under 1% in February 2026." The chart image is the same file in both. No such projection appears in the post's visible text.

Validation

To check the judge, Anthropic asked staff who own the relevant work areas to rate the relative automation of their areas without seeing the models' evidence or ratings ("What this does and doesn't capture"). The post reports exact agreement of 59% between model and human and 35% between humans, and model and human ratings within one level of each other 97% of the time. It summarizes this as the judge agreeing with humans "about as often as humans agreed with each other"; the printed figures put model-human agreement above human-human agreement. The post does not give the number of staff, areas or ratings in the comparison, or the human-human rate within one level. It notes "real room for disagreement on borderline cases", such as the line between AL3 and AL4.

The post also tested whether a frozen basket misses new kinds of work. Building a second tree from January 2026 data and comparing new tasks arriving each month from February to July against it, it found "no rise in the number of 'novel' tasks", which it reads as the structure of model R&D work being stable "at our level of analysis". It plans to rebuild the basket periodically and re-version the published numbers.

Measure 2: oversight of AI agents

Method

The post defines three metrics ("(2) Measuring oversight of AI agents"): coverage, the share of an agent's actions that pass through a monitor before or after execution; review latency, the time from an action to its review by an automated monitor and then a human; and escalation rate, the share of agent activities blocked or redirected (online monitors) or flagged for further review (offline monitors). Online monitors act in real time against fast, irreversible harms such as copying model weights out; offline monitors read activity afterwards for slower, reversible problems such as "signs of model misalignment".

Results

As of August 2026, about 30,000 agents were doing research and engineering work at Anthropic at any one time on its most-used internal platform, and the measurements cover that platform only ("What we found"). The post's table reports:

  • Online monitor. 100% of these agents' actions pass through it before execution, usually within seconds; humans review blocked actions within one week; of over a billion decisions in August 2026, 0.002% (about 1 in 47,000) were blocked.
  • Offline monitor. 100% of actions are ingested after they are taken; it flags roughly one to two transcripts per thousand, about 100,000 a week, which multiple classifiers review for false positives; about 50 a week of the highest priority are escalated to humans.

The post states that "In our monitoring data to date, individual agents rarely misbehave." It reports no rate at which known misbehavior is caught; it recommends that developers share that rate and have third parties test it, and mentions a past METR red-teaming of the offline monitoring platform. The linked Risk Report describes that exercise as one METR staff member spending three weeks red-teaming the automated offline monitoring pipeline in February and March 2026, finding "several specific novel vulnerabilities, some of which have since been patched" (Section 2.23.1.1, p. 77).

The Appendix ("Oversight of agents") describes two design choices in the internal agent scaffold: each agent has an individual identity, independent of the underlying model, to which all its data is tied, and agents communicate through a shared open messaging system cross-referenced with their transcripts. The post presents these as aids to auditing and to agents checking one another; it reports no measurement of their effect. It adds that "We have been doing this kind of monitoring for months, not years".

Comparison with the August 2026 Risk Report

The post links the August 2026 Risk Report as where it "published all of these measurements"; the link (anthropic.com/aug-2026-risk-report) resolves to the redacted PDF (186 pages, coverage date 15 July 2026), the same file the post links for the METR red-teaming. A text search of that PDF, as downloaded on 13 November 2026, found none of the post's figures (30,000 agents, 0.002%, 1 in 47,000, 100,000 flags per week). The Risk Report's Section 2.23.1.1 (pp. 76 to 77) describes offline monitoring of "most internal users" as two monitors: a completion-only monitor with no subsampling for the most common agent scaffolds, and a prompt-and-completion monitor on a 25% subsample of potentially impactful tool calls, with about half of agent-scaffold tokens seen by the latter, and human review of "on the order of 50 trajectories per week". The two documents describe different scopes and dates (one platform in August against most internal users at 15 July), and the post does not say which offline monitor its 100% coverage refers to. The Risk Report also says its setup "does not cover some of" agentic usage, that coverage is "occasionally accidentally" weakened, and that some non-coding use is excluded (pp. 76 to 77); those statements concern most internal users, not the one platform the post measures. The difference is noted here, not resolved.

Measure 3: compute spent on safety

Method

Anthropic took a snapshot of all its compute from 13 to 20 July (2026 by context), stitching together existing capacity-planning tools that tag workloads with best-effort labels, plus usage reported by third-party cloud providers ("(3) Measuring compute allocation" and Appendix). A prompted Claude classifier then labeled each workload as safety work or AI R&D. Safety work is "work whose dominant purpose is making AI systems safer, more understandable, or more secure"; everything else, including work that helps capability as much as safety, counts as AI R&D, which the post calls conservative. For research training and evaluation runs, the classifier read run metadata and code, and about 14% of the week's almost 10,000 runs were sampled, weighted toward the largest by compute. For inference by AI research agents, a variant read session transcripts; where transcripts were inaccessible, usually because work was compartmentalized, the workload was classified by the user's team or defaulted to AI R&D. Safeguards classifiers, which the post says use "a separate, comparable amount of compute", are excluded.

Results

Over the week, about 6% of compute going to AI R&D was allocated to safety, and about 12% of compute going to "AI-driven AI R&D" ("What we found"). The post does not define the second category beyond the Appendix's treatment of inference for research agents. The post itself says compute is "an imperfect proxy" for safety focus, because safety research tends to need fewer compute-intensive runs than frontier training, and that the measure's value lies in comparing like with like over time and across developers.

The classifier prompt and its definitions

The compute classifier is driven by a written definition. The post prints an excerpt of its prompt, with the categories and boundary cases elided: it tells the model it is "helping to perform an internal audit" whose aim is "a public-facing breakdown" of accelerator use, and that its job is to "look at one research job at a time, figure out what it was doing, and assign it to one of those two buckets." The post reports that "an extensive written definition of each task, with clear boundary cases" brought the classifier within "one or two percentage points" of human reviewers, without giving the size of the human-reviewed sample or a comparison with a shorter definition. It also says some cases "were too difficult to determine even after several hours of human review" and that "Our definition is one reasonable choice among many". The prompts for the Automation Index's research agents and judge are not printed.

Proposals and commitments

The post's statements of intent are separate from its measurements:

  • Anthropic plans "to embed independent third-party evaluators from multiple organizations" with access comparable to its internal risk teams, to verify safety practices, report incidents and monitor metrics such as these (introduction).
  • It proposes that any frontier developer publish the three measures regularly under a public methodology, verified by third parties or other developers' models, and suggests the measures could trigger stronger requirements, such as "a fixed testing window" before a new model is used for further AI R&D, or commitments on the share of compute going to safety.
  • It names two obstacles to comparing labs: no common methodology, and that "we're using our own models to evaluate our systems", so the judge may share the errors of the model it checks.
  • It links these proposals to Anthropic's Advanced AI Framework, a policy proposal that includes transparency obligations such as risk reports, and expects the numbers to "shift if there were coordination on pacing the frontier".

Limitations and open questions

The post states these limitations: the automation ratings depend on the judge model; the basket is frozen and may miss new work; the agent-monitoring science is young; the compute labels underneath the classifier are best-effort and unverified, so "usage may be mislabeled"; one week of compute data shows that the measurement can be made but not a trend; and compute share measures only spending, so a more efficient safety classifier lowers the safety share without less safety work being done.

Further facts from the post that bear on how far the numbers go: every measure is produced by Claude models (task listing, tree building, research, judging and classification), and the only reported human validations are the staff comparison for the Index and the reviewer agreement for the compute classifier, each without a sample size; none of the snapshot has been checked by a third party; the oversight figures cover one internal platform; and the publisher is the developer whose pace is being measured, reporting measures it proposes others adopt.

Research questions

  1. How stable is the Automation Index under a different judge? Re-rating the frozen tree with another developer's model, or with independent human raters at scale, would show how much the 26% depends on the Claude judge.
  2. What do the 90% intervals capture? A published account of how the intervals were computed (sampling of staff, of tasks, or of judge runs) would show what uncertainty the chart represents.
  3. Does "leads" work track measured output? Comparing the Index with outcome measures such as time to ship or research throughput over the same months would test whether a rising AL4 share corresponds to faster R&D.
  4. What is the catch rate of the oversight system on known misbehavior? Red-team injections or model organisms run through the online and offline monitors, as the post recommends, would turn coverage figures into a measure of effectiveness.
  5. How sensitive is the 6% safety share to the definition? Re-running the classifier with alternative definitions, or having outside reviewers relabel a random sample of runs and transcripts, would bound the effect of the boundary choices the post concedes.
  6. Do other developers adopt comparable measures? Publications from other labs using the same scale or metrics, and any third-party verification, would show whether the proposal becomes cross-lab practice.

How it connects

  • When AI builds itself (June 2026), the Institute essay the post links on recursive self-improvement, argues that AI is already speeding up Anthropic's work from mostly internal and model-judged evidence; that note records this post as its later quantitative counterpart.
  • Focus areas for The Anthropic Institute (May 2026) asked what telemetry would measure the aggregate speed of AI R&D; this post does not link the agenda, and the connection is this note's.
  • Risk Report: August 2026 (August 2026) is the report the post links for its oversight measurements and its METR red-teaming; that note summarizes the report's monitoring section and recall tables.
  • Natural Language Autoencoders (May 2026) and From shortcuts to sabotage (November 2025) are linked by the post as examples of safety work that uses compute.
  • Measuring AI agent autonomy in practice (February 2026) measures autonomy in external use of Claude rather than inside Anthropic; the post does not link it, and the connection is this note's.

References