Anthropic says Claude now leads a quarter of its own research. It published the yardstick, not just the number.

Anthropic published three internal metrics on 17 September 2026 intended to let outsiders — and other labs — track how fast frontier AI development is actually moving inside a lab, rather than inferring it from product launches. The company frames this as a transparency measure: “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.”

The headline figure is that Claude currently operates at “AL4” or above on 26 percent of Anthropic’s cataloged AI R&D tasks. Anthropic built a six-point scale, AL0 through AL5, where AL0 means no AI involvement and AL4 means the AI “can complete most of the task end-to-end from a high-level prompt, while the human supervises.” The company says it catalogued every category of AI R&D work done internally, rated the current automation level of each, and aggregated the ratings to produce the 26 percent figure.

Why it matters

This is a rare case of a frontier lab publishing a defined methodology for a claim as sensitive as “how much is AI accelerating its own development,” rather than a vague statement. That distinction matters because the number itself is close to meaningless without the scale behind it — AL4 means human-supervised completion, not fully autonomous AL5 work. Readers evaluating lab-pace claims elsewhere in the industry now have a concrete example of what a methodology-backed number looks like, against which vaguer competitor claims can be measured.

What the compute and oversight figures show

Anthropic also disclosed compute-allocation figures from a one-week internal snapshot taken 13 to 20 July 2026: about six percent of compute directed to AI R&D went to safety work directly, and about twelve percent of compute used for AI-driven AI R&D was allocated to safety-focused work. These are vendor-reported, self-disclosed figures from a single one-week window, not an audited or ongoing published series, and Anthropic has not stated whether it will repeat the snapshot on a schedule.

On agent oversight, Anthropic reported roughly 30,000 agents operating simultaneously within the company as of the same reporting period, with about 0.002 percent of agent decisions blocked by monitoring systems and roughly one to two transcripts per thousand flagged for human review. Anthropic said it plans to embed independent third-party evaluators with access to internal processes, systems, and data comparable to what its own internal risk-assessment teams have — a stated intention, not yet an operating programme, and no start date or partner organisation has been named.

It follows 18BYTE’s earlier coverage of Anthropic’s Q2 financial disclosures and its reported compute lease.

What this means for buyers

None of these figures change pricing, availability, or capability of any Anthropic product today — this is a disclosure about internal process, not a launch. Enterprise buyers assessing Anthropic’s AI-safety posture as part of vendor due diligence now have a concrete, if self-reported, artefact to cite: a published automation scale and a named compute-allocation percentage, rather than a marketing claim about “responsible scaling.”

Buyers should treat the specific percentages as a single data point from Anthropic’s own systems rather than an audited baseline, and should ask Anthropic directly whether it intends to publish these metrics on a recurring cadence — a one-off disclosure is a weaker signal of ongoing transparency than a repeated series would be.

What would change our reading

The clearest test of whether this disclosure is substantive rather than a one-time public-relations move is whether Anthropic actually names and onboards independent third-party evaluators with the access it described, and on what timeline. A published second snapshot of the compute-allocation figures — showing whether the six and twelve percent safety shares rise, fall, or hold steady over subsequent quarters — would tell us whether this is a sustained transparency practice. We would also revisit this if a competing lab publishes a comparably specific methodology, which would indicate the industry is converging on this kind of disclosure rather than Anthropic acting alone.