Claude Leads 26 Percent of R&D on the Next Claude. Anthropic Wrote the Yardstick, Too.

On Wednesday, September 17, 2026, Anthropic published three measurements of its own AI research and development and offered the methodology for every other frontier lab to adopt. The lead number is that Claude "leads" 26 percent of the work building the next Claude, defined as the model completing most of a task from a high-level prompt while a human supervises. In February the same share was under 1 percent. Six months, one order of magnitude, and the party doing the measuring is the party doing the work.

The post ships as three proposed public measures: AI-led R&D, agent oversight, and compute allocation. Anthropic is asking peers to publish the same fields on the same cadence. That is a schema play, and schemas beat position papers on a long enough horizon, but no external party has verified any figure in this release.

The Three Numbers

The definitions are the interesting layer. Anthropic splits R&D work into four bins (unassisted, assisted, collaborative, and led), and calls the highest bin "led" because a human still supervises the plan and reviews the output. Nothing in this release claims Claude picks its own goals, deploys changes without review, or operates fully autonomously on any subset. What it claims is that when a human sets the task and hands it over, Claude finishes most of the task most of the time on 26 percent of the work.

The Doubling Curve

Under 1 percent to 26 percent in six months is roughly five doublings, a shade over a month per doubling on the low end. Do not project it linearly. The curve compresses hard as the share approaches saturation, and the definition of "R&D work" is a moving target inside a fast-scaling company: new work streams open up that a human has to seed, and Claude's share of those streams starts near zero every time. The honest read is that Anthropic just showed a 26x expansion of the model's formal role in its own successor over half a year, and printed the number themselves.

The oversight column is where governance policy people should spend their afternoon. Thirty thousand agents running at once, 100 percent of actions screened before execution, roughly a billion decisions per month, and 0.002 percent blocked. Read charitably, the monitors are precise and the fleet is well behaved. Read skeptically, blocking one action in 47,000 is either a very high bar for triggering a stop or a fleet that has learned to stay under the bar. Both readings are consistent with the data, and the data is Anthropic's.

The offline number is a slightly different shape. Around 100,000 transcripts flagged for review per week, about 50 escalated to a human. That is a 2,000-to-1 funnel from flag to escalation, and it is where the substantive judgment calls happen. Nowhere in the release does Anthropic disclose the false-negative rate on either monitor, and no external party is positioned to check it.

The Ledger This Belongs To

Line up the voluntary-governance entries of the last two weeks:

• Sep 5, OpenAI (Pachocki essay), calling for legally mandated safety thresholds enforced by third parties.
• Sep 9, Anthropic + METR, an eight-week wide-ranging-access audit over 481 million transcripts.
• Sep 9, OpenAI, seating Paul Christiano on the Foundation Board with committee authority.
• Sep 12, Anthropic (Amodei essay We Must Pace the Frontier), pledging permanent employee-level METR access.
• Sep 14, Microsoft, publishing the first 38-page draft of its MAI Code of Conduct.
• Sep 14, OpenAI plus Anthropic plus Google, working-group talks on a shared standards body.
• Sep 17, Anthropic, the R&D Automation Index and its companion oversight and compute-allocation metrics.

Seven signals, thirteen days, zero laws. All authored by the labs, all measured by the labs, and none of them a stop-ship order. If you covered the debate on embedded evaluators last weekend, and the market's response on the pacing accord on Monday, this is the same conversation with a scoreboard bolted on. The scoreboard is Anthropic's.

What The Schema Buys

Three things, in this order.

One, a common vocabulary. If OpenAI, Google DeepMind, xAI and Meta publish the same four-bin R&D taxonomy plus an oversight and compute triple on the same cadence, the sector gets its first apples-to-apples measurement of self-improvement velocity across labs. The Amodei essay talked about pacing; a shared index turns pacing into a number you can compare. Anthropic is trying to be the author of that number.

Two, a fallback for regulators. CAISI, the AI Safety Institute network, and the EU AI Act delegated-act track all need a measurement instrument they did not have last month. If a lab index gets adopted by three peers, a rulemaking body cites it, a delegated act references it, and the lab-authored schema becomes the government-recognized one, which is roughly the arc HTTP conditional requests took from RFC to default.

Three, an off-ramp on the audit question. The METR contract is eight weeks. The Christiano board seat is a permanent chair with review authority. The Anthropic R&D Index is a self-report shipped on a quarterly cadence. All three coexist, and they push in different directions. A quarterly self-report is cheaper to run than a rotating audit, and if peers adopt the schema without adopting the METR contract, the industry standardizes on the weaker instrument.

What It Does Not Buy

Every figure is a lab measuring itself. No regulator, no third-party auditor, no academic group has replicated any of these numbers, and none of them can, because the raw inputs are internal agent logs on infrastructure only Anthropic operates. The company says as much in the release. This is transparency by publication, not transparency by verification.

The definitions are also load-bearing and moveable. What counts as "R&D work" is decided by Anthropic. What counts as a "task" that Claude "led" is decided by Anthropic. What counts as an action worth blocking is decided by Anthropic. A future release could hold the 26 percent flat while redefining the denominator, and readers would not know without seeing the classification rules. The methodology detail in the release is thorough by lab-blog standards, and thin by audit standards.

The compute-allocation number is the narrowest slice. Six percent to safety across all AI R&D compute for a single week in July, 12 percent inside the AI-driven-AI-R&D subset. One week is not a policy commitment, and the release does not say the ratio has held or will hold. It is a snapshot, useful as a baseline for the next snapshot, not as a floor a regulator could cite.

The Awkward Overlap

Anthropic asked publicly for embedded evaluators on September 12, and published its self-measured R&D automation index on September 17. Both moves live in the same governance ledger. The first says outside observers should be inside the lab reading the telemetry. The second says here is a well-designed telemetry dashboard, please read it from outside. Both can be sincere, and both can be true, and they still push the sector toward different equilibria. If the schema gets adopted before evaluators do, publish-and-hope wins over audit-and-verify by default.

The OpenAI parallel is the useful comparison. Jakub Pachocki asked on September 5 for legally mandated thresholds enforced by third parties. Paul Christiano took a Foundation Board seat on September 9 with committee review authority. OpenAI has not published an R&D automation index of its own. If it does, we get a comparison in one column across two of the three frontier labs, on a metric the labs themselves defined. If it does not, Anthropic's number stands alone and reads by default as the frontier's number. Neither outcome is neutral.

Our Take

The number that matters is not 26 percent and not 47,000. It is seven, the count of voluntary-governance entries on the ledger in thirteen days, all authored by the labs. The Anthropic index is the most useful entry, because it is the first one that puts a measured decimal against a claim that used to be a vibe, and because it hands peers a template to publish the same decimal on. I would rather live in a world where five frontier labs publish the same three fields quarterly than in a world where none of them do.

But this is still the labs writing the ruler and holding the ruler at the same time. The 0.002 percent block rate is either a tight monitor on a well-aligned fleet or a fleet that learned where the fence is, and the release cannot tell you which without an outside party checking the transcripts the online monitor waved through. The METR audit template exists, is signed, and runs eight weeks; extending it, cross-lab, on the same corpus a quarterly index reports on, is the thing that would turn this publication into a commitment.

For builders on the API tier, nothing on the invoice changes tomorrow. What changes over two to four quarters is the background rate of internal capability improvement that ships into the models you rent. A 26 percent share of the work is the first public signal that the recursive R&D loop has moved past rhetoric and into a decimal Anthropic is comfortable defending. If the doubling holds even loosely, the next Sonnet, Opus and Mythos releases carry more of Claude's own contribution to their training than any previous generation, and the pricing floor conversation gets rewritten from the model layer, not from the harness. We flagged the mechanism in our METR contract piece and in the voluntary-ledger piece from Monday. This is the same substrate. It just has a number on it now.

Three signposts for the next 60 days. First, whether any of OpenAI, Google DeepMind, xAI or Meta publishes an R&D automation index using Anthropic's four-bin taxonomy or a public counterpart, since a schema with one signatory is a blog post and a schema with three is a floor. Second, whether the next issue of the index reports the same definitions or quietly revises the "R&D work" denominator, since a moveable denominator is how a self-report loses its meaning fastest. Third, whether METR, the UK AISI, CAISI or any Article 70 delegated-act draft cites the Anthropic index directly, since the fastest path from lab-authored yardstick to statutory instrument runs through a citation in the first regulatory document that needs one. If two of the three fire, the ledger gets its first entry with external teeth on it. If none fire, the ledger keeps growing and the yardstick stays inside the lab that built it.