Anthropic Says Claude Now Leads 26% of Its Own AI R&D
On this page
Anthropic just published a number I did not expect to see this soon: Claude now “leads” 26% of the work that goes into building the next Claude. Six months ago that figure was under 1%. The company laid it out in a September 17, 2026 post from its policy research arm, the Anthropic Institute, alongside a pile of internal data on how it watches its own AI agents and how much computing power it spends on safety work. I went through the methodology and the raw numbers, and the gap between February and August is the real story here — not the 26% itself.
What “leads” actually means (it’s not autonomy)
Anthropic didn’t just eyeball this. It built something it calls the R&D Automation Index, and the process behind it is more interesting than the headline stat. Researchers sampled 20% of Anthropic staff at random across every department in July 2026 and asked them to catalog the AI research and engineering tasks they actually do. That produced roughly 15,000 granular task descriptions, which the team organized into a tree of 542 nodes, 378 of them leaf tasks with no further children. Every leaf got rated on Epoch AI’s automation scale, AL0 through AL5, and the ratings were weighted by person-time spent, on the theory that a task ten engineers work on full-time should count for more than one an intern touches for an afternoon.
The scale matters because “leads” sits well below “replaces.” Here’s how Anthropic itself draws the line:
- AL3 — “collaborates”: the AI does large chunks of a task, but only under close human direction.
- AL4 — “leads”: the AI can complete most of a task end-to-end starting from a high-level prompt, while a human supervises.
- AL5 — “fully autonomous”: not yet reached at Anthropic, for any measured slice of R&D.
So the 26% figure describes work where a person still has to sign off, not work happening with nobody watching. Zoom out one level and the number gets bigger: more than 90% of Anthropic’s AI research and development now involves Claude at AL3 or above, meaning collaboration with the model is close to the default rather than the exception.
| Metric | Figure |
|---|---|
| Claude “leads” (AL4) share of R&D, Feb 2026 | <1% |
| Claude “leads” (AL4) share of R&D, Aug 2026 | 26% |
| Work at “collaborates” (AL3) or above | >90% |
| Concurrent research/engineering agents (Aug 2026) | ~30,000 |
| Online-monitor block rate | ~1 in 47,000 actions (0.002%) |
| Compute share on safety research (total) | 6% |
| Compute share on safety research (AI-driven R&D only) | 12% |
30,000 agents, and how Anthropic says it’s watching them
The scale numbers are the part that stuck with me. As of August 2026, Anthropic reports roughly 30,000 AI agents doing research and engineering work simultaneously on its main internal platform. Every single action those agents take passes through an online monitor within seconds, before it executes. Across a billion decisions the company analyzed, that live layer blocked about one action in 47,000 — a 0.002% intervention rate.
A second, offline layer reviews 100% of actions after the fact using a stack of classifiers. Out of roughly 100,000 transcripts reviewed weekly, one to two per thousand get flagged for closer review, and around 50 high-priority cases per week get escalated to an actual human. If you want a comparison point on why that oversight layer matters, we covered a case where autonomous AI agents were misused for a real cyberattack in an earlier incident Anthropic disclosed — the monitoring infrastructure described in this report is largely the same machinery built to catch that kind of misuse before it does damage.
Where the compute actually goes
Anthropic also opened up its compute books for a single snapshot week, July 13–20, 2026. A Claude-based classifier sorted about 14% of roughly 10,000 weekly compute jobs, weighted toward the largest ones, to estimate how the company’s total R&D compute budget splits. The result: 6% of all AI R&D compute went to safety research, rising to 12% when you narrow the denominator to just the AI-driven portion of that work. Anthropic flags both numbers as “deliberately conservative estimates,” which is a reasonable hedge given the sampling method, but it’s still a more specific figure than any lab has volunteered before.
Why Anthropic is publishing this at all
None of this happened in a vacuum. Anthropic CEO Dario Amodei spent early September arguing publicly that frontier labs need to slow down and coordinate rather than race, and OpenAI’s policy chief confirmed days later that OpenAI, Anthropic, and Google DeepMind have been holding informal safety talks for weeks — we broke that story in our pacing-the-frontier coverage. This Automation Index reads like the data-backed follow-through on that argument: Anthropic is explicitly asking competitors to publish comparable numbers using the same methodology, covering AI-led R&D share, agent-oversight coverage and escalation rates, and safety compute allocation, ideally with independent verification. The company’s stated reasoning is blunt: “developers, governments, and the wider research community would benefit from converging on a shared definition ahead of time,” rather than after something goes wrong.
Whether OpenAI or Google DeepMind actually take Anthropic up on it is the thing to watch next. A single company publishing its own numbers, audited by no one outside the building, is a start — not proof. But it does put a number on something regulators and researchers have mostly been guessing at for two years: how much of frontier AI development is already being done by AI itself. That number just quietly crossed a quarter, in six months, with no sign of the curve flattening.
Frequently asked questions
What is Anthropic’s R&D Automation Index?
It’s an internal measurement tool that maps roughly 15,000 individual AI research and engineering tasks at Anthropic, rates each one’s automation level on Epoch AI’s AL0–AL5 scale, and combines the ratings (weighted by person-time) into a single aggregate percentage.
Does 26% mean Claude is doing a quarter of Anthropic’s job autonomously?
No. The 26% figure is work at AL4 (“leads”), meaning Claude can complete most of a task end-to-end from a high-level prompt while a human still supervises. Fully autonomous AL5 work has not been reached for any measured task category at Anthropic.
How many AI agents does Anthropic run internally?
About 30,000 agents were running simultaneously on Anthropic’s main research and engineering platform as of August 2026, according to the report.
How much of Anthropic’s compute goes to safety research?
Anthropic estimated 6% of total AI R&D compute and 12% of AI-driven R&D compute specifically went to safety research during a sampled week in July 2026, describing both figures as conservative estimates.
Why does this matter outside of Anthropic?
Anthropic is using the report to push OpenAI, Google DeepMind, and other frontier labs to publish comparable automation, oversight, and safety-compute data using a shared methodology, arguing that a common definition now is better than a patchwork built after an incident forces the issue.
Sofia follows emerging technology, from AI and VR to IoT and blockchain, and translates the hype into plain language. She cares about what these tools mean for everyday users, not just the headlines.
