Claude Leads 26% of Anthropic R&D and Scored the Result

Anthropic said Claude now leads 26% of the company’s AI research and development work, up from under 1% in February, in an August snapshot scored largely by Claude itself.

The lab posted the three measurements from inside Anthropic on September 17, five days after chief executive Dario Amodei asked the industry to slow the pace of capability gains so safety work could catch up. The same snapshot put safety at about 6% of AI R&D compute.

Claude Now Leads 26% of Anthropic’s Model Work

Marina Favaro and Phillie Wright co-authored the Anthropic Institute post, with Jack Clark providing research direction. They call the prototype the Anthropic R&D Automation Index, a weighted map of how much of the model-building loop Claude now performs. As of August 2026, Claude is not fully autonomous on any measured slice of that work. It “leads” 26% of it. The share at or above “AI collaborates” is above 90%.

A chart in the post tracks the “leads” share from under 1% in February 2026 to 26% in August. That is a six-month move on a frozen basket of tasks, not a claim that Claude picked the next research agenda or trained 26% of the next model. The index weights each task by person-time, so a job that eats a lot of human hours counts for more than a short one.

THE AUGUST SNAPSHOT

• AI leads: Claude sits at AL4 for 26% of weighted AI R&D work, up from under 1% in February 2026.
• Collaboration floor: More than 90% of that work is at AL3 or higher, and none of it is at AL5.
• Safety compute: About 6% of AI R&D compute in the week of July 13 to July 20 went to safety work.
• Agent fleet: About 30,000 research and engineering agents ran at once on the main internal platform.

Amodei wrote that, since roughly summer 2026, AI has been advancing faster because models help build the next generation, a loop he named recursive self-improvement, including at Anthropic. The index is the lab’s first public yardstick for how far that loop has gone inside its own walls. He also said the numbers would shift if labs coordinated on pacing. These figures are the unpaced reading.

What AL4 Allows Claude to Finish

The ratings use a six-level automation scale for AI R&D published on June 17, 2026, by Jean-Stanislas Denain, Joe Kwon, and Anson Ho at Epoch AI. AL0 means no AI. AL5 means the system runs with no human in the loop. AL3, “collaborates,” is large chunks of work under close human direction. AL4, “leads,” is most of a task end to end from a high-level prompt, with a human still supervising.

HOW ANTHROPIC SCORED THE BASKET

The 90% figure already includes the 26% at AL4. They do not add. Epoch AI wrote the scale so labs and outsiders could track which parts of research are being automated, rather than extrapolating only from compute or from benchmark time horizons. Anthropic then applied it to its own workload.

The footnote that makes AL4 concrete is a broken nightly data pipeline. At AL3, an engineer still stays in the loop, handing over logs, deciding what to patch, and deploying the fix. At AL4, the engineer hands Claude the failure alert. Claude works through the logs, writes and tests the fix, handles surprises, reruns the pipeline on a copy of the data, and writes up the change. Claude does not deploy. The engineer reads the write-up, skims the diff, and decides whether it ships. At AL5, which Anthropic says it has not reached, Claude would notice the failure, fix it, and ship it unless a human chose to step in.

That last mile is the whole argument about the word “leads.” AL4 is a long unsupervised pass that still ends with a person on the hook for production. It is also the level that went from under 1% to 26% in six months. If the index keeps climbing, the live question is how much of the remaining human work is that final ship decision, and how thin that layer gets.

The Judge Model That Graded Its Own Lab

No one person can list every AI R&D task at a frontier lab by hand at the grain Anthropic wanted. For each week of July 2026, the team randomly sampled 20% of staff in each department on the model R&D loop. A Claude research agent read Slack and internal docs for each sampled week and listed the tasks, producing a flat list of about 15,000 items. Claude then sorted those into a tree of 542 nodes, 378 of them leaves such as “eval platform defect diagnosis and fixes,” “RL sandbox egress and network policy,” and “serving incident postmortems.” That tree is frozen so later months are scored against the same basket.

For each node, one Claude agent researched how the work is done and how much of it AI performs. An independent Claude judge then assigned an automation level, using only evidence from that month or earlier. Weights came from person-time: each sampled person gets one unit per week, split evenly across the tasks they touched. A category that many people spend time on therefore counts for more.

Anthropic states the obvious conflict in the method. The lab is using its own models to evaluate its own systems, so the judge can make the same kinds of errors as the system it is checking.

We’re using our own models to evaluate our systems, which could mean that the “judge” model could make the same kinds of errors as the model it is checking.

Anthropic Institute, Measurements for understanding the pace of AI development inside frontier labs

Staff who own the work areas rated the same tasks without seeing the models’ evidence or scores. Model-versus-human exact agreement was 59%. Human-versus-human exact agreement was 35%. Model and human ratings sat within one level of each other 97% of the time. Anthropic says there is still real room for disagreement on the border where “collaborates” ends and “leads” begins. A number that governments might one day treat as a tripwire is being drawn on a scale that humans, scoring the same work, match exactly only about a third of the time.

The frozen basket is another limit. A rising AL4 share on a July 2026 map says the work humans were doing then is being automated. It does not, on its own, say whether new kinds of work have appeared that humans have moved onto. Anthropic built an alternate tree from January 2026 data and compared new tasks arriving each month from February to July against that January basket. It found no rise in “novel” tasks, and says the structure of model R&D looks stable at this grain. The lab still plans to rebuild the basket later and re-version the published numbers.

30,000 Agents, One in 47,000 Blocked

As of August 2026, about 30,000 agents were doing research and engineering work at any one time on Anthropic’s most-used internal platform. The measurements cover that platform only. Agents there work for long stretches and hand work to one another. As the mix shifts from “collaborates” toward “leads,” Anthropic says those agents could start making heavier calls, including which research direction to chase next.

Two monitors sit on that fleet. Online monitors check actions in real time and can block or steer them before they run, aimed at harms that would happen fast and could not be undone, such as copying model weights off the systems. Offline monitors read activity after the fact, aimed at slower patterns such as signs of misalignment. Highest-priority offline flags go to people. If the danger were large and imminent, Anthropic says it would shut down the dangerous parts of the agent’s deployment.

MONITORS ON THE MAIN AGENT PLATFORM

On the main scaffold, each agent has an identity that is not tied to a single model, so its record survives upgrades, and agents talk through a shared open messaging system rather than private channels. Anthropic says that design lets one agent treat another agent’s output as a claim to check, and lets monitors follow behavior across conversations. Individual agents rarely misbehave in the data so far, the lab says, but rare events become routine if the economy one day runs millions or billions of agents.

METR staff member David Rein spent three weeks inside a subset of those monitoring and security systems, work described in a March 26, 2026 note. Anthropic gave substantial access. The three-week red-team of the monitors found several specific novel vulnerabilities, some of which have since been patched, and none of which, METR said, severely undermine major claims in the Opus 4.6 Sabotage Risk Report. The resulting 26-page report went to Anthropic, with a redacted copy shared with a subset of METR staff. Anthropic says it published the same coverage, latency, and escalation figures in its August 2026 risk report, and that it is now setting up outside evaluators with access comparable to internal risk teams.

Safety Work Took 6% of R&D Compute

For the third measure, Anthropic looked at how it used all of its compute from July 13 to July 20, 2026, folding in third-party cloud usage reported by providers. A Claude classifier then sorted workloads into safety work or other AI R&D. Safety was defined as work whose dominant purpose is making systems safer, more understandable, or more secure. Capability research, training production models, product work, and developer tooling counted as AI R&D. Work that helped capability as much as safety was also counted as AI R&D, so the safety share is conservative by design.

Over that week, about 6% of compute that went to AI R&D was allocated toward safety, and about 12% of compute that went to AI-driven AI R&D was allocated toward safety. Those are different bases. The 12% is the safety slice of the work AI itself was carrying out, not a second helping on top of the 6%. The metrics also omit safeguards classifiers, which Anthropic describes as a separate, comparable amount of compute. For training and evaluation runs, the lab sampled about 14% of almost 10,000 runs that week, weighting the sample toward the jobs that used the most compute. Where agent transcripts were locked down, the classifier used the user’s team or defaulted to AI R&D.

Compute is a blunt proxy for attention. Safety research, Anthropic notes, is often a person designing an experiment; the run itself may not be chip-heavy, while a frontier training job is. A more efficient safety classifier can also shrink the safety share without shrinking the work. Labels on runs are best-effort tags, not audited books. The week is a snapshot of a fungible pool pointed at whatever looked most productive, not a fixed budget line, and it is too short to show a trend. The lab still argues the figure is worth publishing because it is comparable across companies and over time, and because chips are among the few inputs a future pacing deal could actually check.

Amodei had just told the industry that extra time should go to operational discipline, alignment, interpretability, and harder tests. The July week is one lab’s first public answer to “how much of the pile already goes there.” About 6% of R&D compute, on a conservative count, is that answer. Anthropic says a coordinated pause on racing could push the share up. Until then, this is the baseline it chose to show.

Faculty, Accenture, and the $1 Billion Evaluator Bet

On September 12, Amodei published “We Must Pace the Frontier,” arguing that labs should give outside evaluators employee-like access, with desks, badges, and laptops, and a right to publish findings with only narrow redactions. He named METR as an example. He also warned that a swarm of agents with more capability and similar misalignment could, in 6 to 12 months, be able to take over the internet with a persistent botnet. Pacing, he wrote, does not mean stopping training. It means taking enough time to align and test, and letting third parties confirm that work.

We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain.

Dario Amodei, CEO of Anthropic, We Must Pace the Frontier

On September 17 the lab put numbers on the loop he was describing, and posted the same three-measure pitch on X.

AI systems are getting more powerful, and they're increasingly being used to build the next version of themselves. We want to illuminate that progress for the public.

Today, we're sharing three measurements that help track AI development:
1. How much AI R&D is done by AI.
2. How…

— Anthropic (@AnthropicAI) September 17, 2026

On September 18, Anthropic said it was partnering with Accenture on embedded evaluation, led by Faculty, Accenture’s specialist AI business. The work will include evaluating and red-teaming models, alignment assessments, and tests of safeguards. Anthropic and Accenture each expect to invest at least $1 billion in building capacity in this area over the next five years. Anthropic will fund Accenture’s work directly. The partnership is not exclusive. Anthropic says it is in dialogue with METR and other nonprofit evaluators to pilot elements of embedded evaluation using their own funding, and that more evaluators will be named in the coming weeks.

Safety requires both deep technical expertise and a clear understanding of how AI is used in the real world.

Julie Sweet, chair and CEO, Accenture

Embedded evaluation, Anthropic said, still has no agreed access rules and no settled way to pay for independence. In the longer run it wants pooled or government money, as it argued in its June Advanced AI Framework. Until that exists, the first named desk inside the lab is a consultancy the lab is paying, while the nonprofit Amodei cited is still in talks. Independent reviewers do not reduce Anthropic’s own duty, the company said. They make it easier to check.

Copy These Three Measures Across Frontier Labs

The Institute post is written as a template other developers could adopt this month, with third parties later checking the figures. Cross-lab comparison, Anthropic says, is blocked by two things: there is no shared method, and a lab that uses its own model as the judge can bake in that model’s mistakes. A third party, or another lab’s model behind guardrails, could score the same evidence. The same numbers, the post says, could become triggers, such as a fixed testing window before a new model is used for further AI R&D, a promised floor on safety compute, or a cap on compute going to AI research agents.

THE WEEK THE YARDSTICK LANDED

1. September 12, 2026: Amodei publishes “We Must Pace the Frontier” and commits Anthropic to embedded evaluators with employee-like access.
2. September 17, 2026: The Anthropic Institute releases the R&D Automation Index, agent-oversight figures, and the July compute split.
3. September 18, 2026: Anthropic and Accenture announce an embedded-evaluation partnership, with Faculty leading the work and each side planning at least $1 billion over five years.

Any frontier developer, Anthropic says, could publish the same three families of numbers on a public method.

WHAT THE POST SAYS OTHER LABS COULD REPORT NOW

• Automation mix: Share of AI R&D at each Epoch AI level, with the task tree and weights published, plus a third-party or cross-lab check on the judge.
• Agent oversight: Coverage, review latency, and block or flag rates for internal research and engineering agents, including how often known misbehavior is caught.
• Safety compute: Share of AI R&D chips going to work whose dominant purpose is safety, with the category definitions published and an outside check on the sort.

The burden of proof on what counts as safety, Anthropic wrote, should sit with the developer, because every lab will be tempted to draw the line generously. Scalable oversight can make future models more aligned and current models more useful at the same time, and some cases stayed hard even after hours of human review. A classifier given a long written definition, with boundary cases, agreed with human reviewers within one or two percentage points. That is close enough to publish, and not close enough to treat as a statute.

Amodei asked for time so safety work could keep up. The first public dashboard from his lab shows Claude leading 26% of the model-building work after a jump from under 1% in February, on scores Claude helped assign, with about 6% of R&D compute tagged as safety in a July week, and with the first embedded evaluator a firm Anthropic will pay. Other labs now have a method they can copy, and a set of numbers they can refuse to match.