Skip to main content

Claude Now Does 26% of Anthropic's R&D—And That's a Warning Shot

Anthropic Puts a Number on AI Self-Improvement

On September 17, Anthropic published a blog post that did something unusual: it quantified how much of its own AI research is now done by AI. The headline figure—26% of internal R&D work is led by Claude—might sound modest until you consider it was under 1% just six months earlier.

At any given moment, roughly 30,000 AI agents are running on Anthropic's internal platform. In August alone, they made over 1 billion decisions. This isn't a demo; it's the daily workflow.

From 'Fix It' to Full Pipeline Repair

Anthropic uses a six-level automation framework from Epoch AI, where AL0 means no AI involvement and AL5 means full autonomy. As of August, 26% of R&D tasks hit AL4—humans set high-level goals, and Claude handles the rest end-to-end. More than 90% reached at least AL3, where AI collaborates with humans.

A late-night data pipeline failure illustrates the shift. Engineers simply told Claude: "fix it." The AI checked logs, found the root cause, wrote repair code, tested it, handled new errors, reran the pipeline, compared results, and wrote an explanation. No human hand-coding required.

But Anthropic is careful to note: none of the measured work has reached AL5, the fully autonomous level.

The Safety Net Behind the Numbers

All agent actions pass through online monitoring before execution. Of the 1 billion decisions in August, about 0.002% were blocked. Offline monitoring flags roughly 100,000 conversation records weekly; after filtering false positives, only about 50 get escalated to human review.

When it comes to compute, about 6% of AI R&D power went to safety research during a sample week in July. If you count only AI-driven R&D, that share rises to 12%. Anthropic admits the line between "AI collaboration" and "AI-driven" is subjective—cross-lab comparisons will need third-party validation.

The RSI Race Heats Up

Recursive self-improvement (RSI) is the hottest phrase in AI right now. Zhipu's Tang Jie said yesterday that GLM-5.3-Flash was fully deployed on domestic chips in two weeks, mostly by an Infra Agent. He thinks full RSI is still far off, but the cycle of "models optimizing systems and serving models" has begun.

OpenAI announced in early September that it hit a milestone with automated AI research interns, aiming for automated AI researchers by March 2028. Google DeepMind's AlphaEvolve already contributed circuit designs to next-gen TPUs—what Jeff Dean called "the TPU brain helping design the next TPU body."

As Google DeepMind's strategy chief Jagjeet Saini put it: RSI is becoming the core investment logic for AI capex. Current revenue can't justify the spending, but not betting on RSI is unwise.

Why Anthropic Showed Its Hand

By releasing these numbers, Anthropic is doing two things. First, it invites society to help steer AI self-iteration while humans still hold the wheel. Second, it sets a benchmark: once you disclose R&D automation rates, agent scale, compute allocation, and safety coverage, transparency itself becomes competitive pressure. Not following suit might look like hiding something.

Key Points

  • 26% of Anthropic's R&D is now led by Claude, up from <1% in February.
  • 30,000 agents make 1 billion decisions monthly; 0.002% blocked by safety monitors.
  • No work has reached full autonomy (AL5) yet, but 90%+ is at least AI-collaborative.
  • 6% of R&D compute goes to safety; 12% if counting only AI-driven work.
  • Rivals OpenAI, Google DeepMind, and Zhipu are all racing toward automated research—but each has different milestones.