Anthropic reveals Claude now drives 26% of its AI R&D, up from under 1% in February
APPSO · wechat · 2026-09-18
Anthropic has published its first detailed look at how much of its AI research Claude already does.
- R&D automation: Using EpochAI's AL0-AL5 scale, Claude "leads" (AL4) about 26% of R&D work as of August, up from under 1% in February; over 90% of tasks reach at least "AI-collaboration" level. Nothing has hit AL5 (fully autonomous deploy). The index was built by having Claude mine Slack/docs into 15,000 tasks across a 542-node taxonomy; full human-model agreement is only 59%.
- Agent monitoring: 30,000 agents run on its busiest internal platform, making over 1 billion decisions in August, with just 0.002% blocked by real-time monitors (1 per 47,000). Post-hoc monitoring flags 100k conversations weekly, 50 go to human review. Agents have persistent identities for full traceability.
- Compute allocation: 6% of AI R&D compute goes to safety research (12% of AI-driven R&D compute) — a conservative estimate, per Anthropic.
- Caveats: Self-assessed by Claude, no cross-lab standard, one week of compute data. Anthropic plans to open internal systems to independent third-party evaluators. Recursive self-improvement hasn't arrived, but its early loop is now visible in real R&D pipelines.
Related event: Anthropic Discloses Internal Metrics: Claude Now Drives 26% of Its R&D(41 posts)→
More from AGI Musings
- UK's 'AI sovereignty' push undermined by its own energy and copyright policy — NinaDSchick · 2026-11-02
- US firms out-extracted Europe on IT via management — same may hold for AI — carlbfrey · 2026-11-02
- Detecting covert instruction-following via internal representations parallels J-space work — dioscuri · 2026-11-02
- Paras Chopra: Mastering AI means building self-improving systems, not solving problems directly — paraschopra · 2026-11-02
- Databases to distributed systems: every systems skill is merging into AI inference and hardware — Abhishekcur · 2026-11-02
- Limitless analogy: smarter people may gain far more from superintelligence — McDonaghMatthew · 2026-11-02