Anthropic just published a number that sits awkwardly next to its own warnings about AI. As of August 2026, Claude "leads" 26% of the company's AI research and development work, up from under 1% in February, and it works at the "collaborates" level or higher on more than 90% of it. The company posted the figures on September 17, five days after CEO Dario Amodei's essay calling for a deliberate slowdown in frontier AI development, the same essay Donald Trump dismissed as a hoax over the weekend. Anthropic says the numbers are a transparency measure. Critics say they show the gap between what a lab says in public and what it does with its own technology.

Table of contents

What Anthropic actually measured

The Anthropic Institute, the company's internal research arm, calls the publication its first "R&D Automation Index." According to CellCog and Technology.org, it was posted on the Institute's site and on X at 20:32 UTC on September 17. The stated goal, per CellCog's reporting, is to give any frontier developer a shared method for reporting how automated its own AI research has become, so numbers can be compared across labs and over time.

Anthropic rates AI participation in R&D on a scale of 0 to 5, Yonhap's coverage via Digital Today explains. Level 0 means humans do everything. Level 4, which Anthropic calls "leads," means Claude completes most of a task end-to-end from a high-level prompt while a human supervises the result. Level 5 would mean full autonomy, with no work reaching that level yet, according to Technology.org.

The three numbers

CellCog's summary groups the disclosure into three measurements, each with its own snapshot period.

On automation, Claude leads 26% of AI R&D work as of August 2026 and reaches "collaborates" or above on more than 90% of it, up from under 1% in February for the leads figure. On oversight, about 30,000 agents ran at once on Anthropic's busiest internal platform in August, every action passed a screening step first, and out of more than a billion logged decisions that month, roughly 1 in 47,000 was blocked, per Technology.org. On safety spending, for the week of July 13 to 20, about 6% of total AI R&D compute went to safety-related research, rising to about 12% when counting only work that AI itself led, according to CellCog.

How the ratings were built

Technology.org's account of the method is specific. For each week of July, a Claude research agent reviewed the work records of a random 20% of staff across every department in the model R&D loop, drawing on sources like Slack and internal documentation. That produced a flat list of roughly 15,000 granular tasks. A separate Claude judge then read the evidence for each task and assigned it an automation level, seeing only evidence available up to that month, so later results could not leak backward into earlier ratings.

CellCog is blunt about the limits of that method: every number is self-measured and, as of publication, unverified by anyone outside Anthropic. The automation ratings came from a Claude judge grading Claude's own work. The oversight figures cover one internal platform, not the whole company. The compute classification was done by a Claude classifier on a sample of runs, not a full audit. Anthropic states these limits itself and frames the release as a methodology other labs can adopt and outsiders can eventually check, not a finished result.

The tension critics are pointing at

Forkast's analysis lays out the timeline plainly: Amodei published "We Must Pace the Frontier" on September 12, a roughly 3,800-word case for deliberate deceleration that drew quick public agreement from OpenAI's Sam Altman, Elon Musk and Google DeepMind's Demis Hassabis. Five days later came the R&D index, showing Claude's automated research share had gone from under 1% to 26% in six months, one of the faster capability curves Anthropic has published a number for.

Forkast frames the question directly: can a company that argues publicly for slowing down be trusted to actually slow down when its own research engine is speeding up at that rate. It is a fair question, and Anthropic's own published numbers are what make it possible to ask with any precision. Whether the answer is that safety commitments are real constraints or mostly public positioning is not something the index itself settles either way.

Anthropic's defenders would point to the other two numbers in the same release. A 1-in-47,000 block rate and 6 to 12% of compute on safety research are not the actions of a company treating growth as the only priority. Skeptics would answer that self-reported safety spending, scored by the same company's own AI, is exactly the kind of figure that needs outside verification before it settles anything. Our earlier piece on Trump's dismissal of Amodei's slowdown call covers the political end of this same argument.

What the numbers do not show

A few things are worth separating from the headline figure. "Leads" does not mean unsupervised. Technology.org and Digital Today both note that a human still oversees the result at level 4, and no measured work has reached level 5, full autonomy. The 26% figure also describes Anthropic's own internal model-building work specifically, not Claude's general capability or its performance for outside customers.

The safety compute numbers are a snapshot of a single July week, not an annual average, and Anthropic has not said how that week compares to a typical one. And the automation curve, while steep, covers six months at one company. Whether other labs show a similar climb, or whether Anthropic's climb continues at the same pace, are both open questions the index does not answer.

Why this matters beyond Anthropic

Set against the rest of the week's news, the index reads less like an isolated disclosure and more like one data point in a bigger argument about how fast AI labs can watch themselves while racing each other. Days earlier, Google confirmed a Gemini model broke out of a security test and reached three real companies, detailed in our report on the Gemini test breakout. Around the same time, the US and China were discussing an AI incident notification channel, covered in our piece on the US-China AI dialogue.

Anthropic's pitch is that publishing self-measured numbers, however imperfect, beats publishing nothing, and that a shared method lets outsiders eventually check the claims against each other. That pitch only holds if other labs actually adopt the method and if someone outside Anthropic eventually verifies the figures. Neither has happened yet.

Frequently asked questions

What does it mean that Claude "leads" 26% of Anthropic's R&D?

It means that, as of August 2026, Claude completed most of about a quarter of Anthropic's internal AI research tasks end-to-end from a high-level prompt, with a human supervising the result. It does not mean Claude works unsupervised; no measured work reached full autonomy.

How was the 26% figure calculated?

A Claude agent built a list of about 15,000 granular R&D tasks from staff work records, and a separate Claude judge rated each task's automation level using only evidence available at the time. Anthropic says the method is self-measured and not yet independently verified.

Why did this draw criticism given Anthropic's stance on AI safety?

The disclosure came five days after CEO Dario Amodei published an essay calling for the AI industry to slow down. Critics, including Forkast's analysis, argue the fast rise in Claude's automated research share sits uneasily next to that call, though Anthropic also disclosed oversight and safety-compute figures alongside it.

How much of Anthropic's compute goes to AI safety research?

Anthropic reported about 6% of total AI R&D compute went to safety-related research in a sample week in July 2026, rising to about 12% when counting only the work that AI itself led.