Short answer: on September 17, 2026, Anthropic published three new public metrics meant to give outsiders visibility into the pace of its own frontier AI development — how much of its AI research is now performed by AI itself, how closely it monitors autonomous AI agents, and what share of its computing resources goes to safety work rather than capability work. Last verified: September 20, 2026. The release, “Measurements for understanding the pace of AI development inside frontier labs,” is primary-sourced to Anthropic’s own newsroom; nothing in this article beyond Anthropic’s self-reported numbers has been independently verified by an outside party, and CASRAI has not independently verified them either.
What Anthropic published, in atomic facts
- Source: Anthropic, “Measurements for understanding the pace of AI development inside frontier labs,” published September 17, 2026, authored by Marina Favaro and Phillie Wright.
- Metric 1 — AI-led AI R&D: the share of Anthropic’s own AI research and development work performed by AI systems, scored on a five-level automation scale (AL0 through AL5).
- Metric 2 — oversight of AI agents: coverage, review latency, and escalation rate for monitoring what autonomous Claude agents actually do.
- Metric 3 — compute allocation: the percentage of Anthropic’s total AI R&D compute spent on safety work specifically, versus other R&D.
- Headline automation figure: Claude now “leads” 26% of Anthropic’s AI R&D work, up from under 1% in February 2026, by Anthropic’s own account.
- Agent-monitoring figures: roughly 30,000 agents active on Anthropic’s primary internal platform, with 0.002% of agent actions blocked by online monitors.
- Compute figures: 6% of total AI R&D compute, and 12% of AI-driven AI R&D compute specifically, allocated to safety work.
- Third-party evaluator plan: Anthropic says it plans to “embed independent third-party evaluators from multiple organizations at Anthropic, and give them access to internal processes, systems, and data comparable to what internal risk assessment teams have,” citing METR’s prior independent red-teaming of its offline monitoring platform as precedent.
Why Anthropic says it’s doing this
Anthropic frames the release around a visibility gap, not a pacing commitment. Its own stated rationale: “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.” That is a carefully hedged sentence — it acknowledges a live public debate about pacing without Anthropic itself committing to slow down, and it frames the three metrics as an input to that debate rather than a policy position.
All the numbers above are self-reported. Anthropic’s own post is explicit that no outside party has yet verified them: the “third-party evaluator” arrangement described above is a stated plan, not something already in place, and Anthropic points to METR’s earlier red-teaming of a different system (its offline monitoring platform) as the closest existing precedent — not as evidence these specific pace metrics have already been checked.
Context: a wider “slowdown” debate this release sits inside
This release lands inside a broader public argument about whether frontier AI development should slow down — a debate that includes Anthropic CEO Dario Amodei’s own past calls for a more deliberate pace and, according to press and opinion coverage, statements attributed to Google DeepMind CEO Demis Hassabis and op-eds from policy institutes including Carnegie, Brookings, and the Atlantic Council, alongside pushback from Chinese state-affiliated media characterizing the pacing argument as geopolitical maneuvering. CASRAI has not independently verified those individual claims against primary sources in this pass — WebSearch capacity for that verification work was unavailable at drafting time — and none of them should be read here as independently confirmed fact. They are noted only as attributed context for why this specific, primary-sourced release landed where it did. A dedicated piece examining that wider slowdown debate on its own is a separate task CASRAI is holding until it can verify each claim directly against primary sources.
Where this fits in CASRAI’s NIKOLAI project
NIKOLAI is CASRAI’s own frontier-AI-safety dictionary of elements — an independent, unendorsed reference work, not a standard any lab, evaluator, or regulator has adopted or confirmed. Its N8 track, “Transparency and review,” exists specifically to catalogue this kind of voluntary lab disclosure. Two N8 elements are directly relevant here, each verified against its live element page before being quoted: External Review, defined as “an assessment performed by a party outside the model developer — of a model, a risk report, a safeguard, or compliance with a framework — capturing the review’s type, scope, and output”; and Publication Rights Clause, which records “the contractual or policy term(s) governing what an external evaluator may publish about a review — who may redact content, and whether the developer retains editorial control or approval rights over the evaluator’s findings.”
Anthropic’s September 17 release is not itself an instance of either element — the three metrics are Anthropic’s own self-reported numbers, not the output of an outside assessment, so there is no External Review record to point to yet. What makes the release a genuine, concrete N8 example is the plan it describes: Anthropic says it intends to bring in independent third-party evaluators with internal-team-level access specifically so reporting like this can eventually be externally checked. If and when that happens, the resulting assessment is exactly the record type NIKOLAI’s External Review element is built to describe, and the Publication Rights Clause element is exactly the question that would need answering to know whether that evaluator could publish an unflattering finding on its own terms, or only with Anthropic’s sign-off. As with every NIKOLAI crosswalk row, this is CASRAI’s own independent, unendorsed reading of a public source — not a mapping Anthropic has confirmed, adopted, or endorsed.
Frequently Asked Questions
What three metrics did Anthropic publish on September 17, 2026?
AI-led AI R&D (how much of Anthropic’s own AI research is performed by AI, on an AL0-AL5 automation scale), oversight of AI agents (coverage, review latency, and escalation rate for monitoring autonomous agent actions), and compute allocation to safety work versus other R&D.
Is this the same as Anthropic committing to slow down frontier AI development?
No. Anthropic frames the release as a transparency measurement exercise aimed at narrowing the gap between what labs know internally and what the public knows — not as a pledge to slow its own development pace.
Has an outside party independently verified these numbers?
Not yet, by Anthropic’s own account. The post describes a plan to embed independent third-party evaluators with internal-team-level access, and cites METR’s earlier red-teaming of a different Anthropic system as precedent, but the pace metrics themselves are currently self-reported.
What is NIKOLAI’s connection to this release?
NIKOLAI’s N8 “Transparency and review” track includes an External Review element (for recording an outside party’s assessment) and a Publication Rights Clause element (for recording what that outside party is allowed to publish). Anthropic’s release isn’t an instance of either yet, but its stated plan to bring in third-party evaluators is exactly the kind of development those two elements exist to track.
Where can I read Anthropic’s original announcement?
Directly on Anthropic’s own site: “Measurements for understanding the pace of AI development inside frontier labs,” published September 17, 2026.
Related reading
- Anthropic and Accenture Commit $1B Each to Scale AI Model Evaluation
- A Cross-Lab AI Safety Standards Body? What’s Confirmed So Far About the Anthropic-OpenAI-Google Talks
- External Review, Unpacked: Which Frontier Labs Let Outsiders Audit Their Safety Cases
- Third-Party Evaluator Standards and Methodology
- What Is NIKOLAI? CASRAI’s Frontier-AI-Safety Dictionary Explained
