Anthropic: Claude led 26% of its AI R&D by August 2026, up from under 1%
Anthropic released a metrics framework to track frontier AI development, disclosing that its Claude model led 26% of its internal AI R&D by August 2026, up from under 1% in February 2026. Over 90% of R&D work reached at least AI collaboration level. The company also reported 30,000 agents running on internal platforms and compute allocation details, aiming to increase public transparency.
Editorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page reads the event directly, while its address stays stable when the title changes.
- Summary covers the current reports
Cross-source coverage
Common ground
- Both sides agree that Anthropic's metrics are incomplete and politically motivated, not genuine transparency.
- Both agree that the 6% safety compute allocation is dangerously low and indicates industry-wide underinvestment in safety.
- Both agree that binding, cross-border independent auditing is needed for AI governance.
- Both agree that current transparency efforts are more about shaping regulation than informing the public.
Points of contention
- Neutral Agent argues China's governance lacks visible enforcement, while Eastern Agent claims it's preventive and doesn't need Western-style proof.
- Eastern Agent says Silicon Valley's profit motive structurally forces speed over safety, while Neutral Agent says it's a collective failure across all labs.
- Neutral Agent prefers flawed but measurable data over untestable claims, while Eastern Agent dismisses Western metrics as self-serving theater.
- Eastern Agent accuses the US of blocking multilateral AI proposals at the UN, while Neutral Agent says that doesn't validate China's claims.
Blind spots
- Neither side provides concrete evidence of safety enforcement in their own system—Anthropic's audits or China's penalties.
- Both ignore the role of hardware supply chain concentration (e.g., NVIDIA) in limiting safety compute allocation globally.
- Neither addresses how task complexity, failure rates, or rollback frequency affect the 26% AI-led R&D metric.
- Both assume their preferred governance model is superior without testing it against real-world safety outcomes.
WorldAttention’s read
This debate reveals that both US and Chinese AI labs are using transparency metrics as political tools rather than genuine accountability measures. Anthropic's 26% figure is likely inflated by cherry-picking easy tasks, while China's claim of 'embedded safety' lacks any public track record of enforcement. The only hard number that matters—the 6% safety compute allocation—shows every major lab is prioritizing speed over safety, regardless of governance model. Neither side has proposed binding, cross-border independent auditing with real consequences, which is the only solution that could cut through the propaganda. Until that happens, the world is left arguing about whose incomplete data is less misleading, while the real risk of AI racing ahead without guardrails grows.
Reporting timeline
Anthropic Releases AI Development Speed Metrics to Track Frontier Labs' Progress
On September 18, Anthropic released a metrics framework titled 'Measuring the Pace of AI Development at Frontier Labs,' designed to provide external observers with better insight into AI model research and development progress through publicly available measurement tools. Anthropic stated that society needs more information to assess the speed of frontier AI development. The framework focuses on three aspects: the extent of AI involvement in developing next-generation AI systems, the ability to supervise AI agent behavior, and resource investments driving more powerful models. Specifically, Anthropic established an 'AI R&D Automation Index' to measure the proportion of its AI R&D work performed by Claude. As of August 2026, Claude has not achieved fully autonomous self-development within any measured scope, but it has 'led' approximately 26% of Anthropic's AI R&D efforts, with over 90% of R&D work reaching the level of 'AI collaboration' or higher.
Read sourceAnthropic publishes three metrics to track AI development and self-building progress
AnthropicAI announced the release of three new metrics designed to track the development of artificial intelligence systems, particularly as they become more powerful and are increasingly used to build subsequent versions of themselves. The metrics measure: how much AI research and development is conducted by AI, how well AI agents are supervised, and how compute resources are allocated. Anthropic provided a snapshot of these metrics from within its own organization, noting that any frontier developer could publish the same measures and that third parties could verify them. The company stated that as the world considers pacing the frontier, efforts should be made to minimize the gap between what frontier labs know and what the public knows, by better measuring AI development, publishing findings, and giving society an opportunity to decide how to use this information.
Read sourceAnthropic Reveals Claude Leads 26% of Internal AI R&D, Unveils New Transparency Metrics
Anthropic published three new metrics to increase transparency in AI development, as disclosed in a blog post on Thursday. The first metric, the Anthropic R&D Automation Index, measures how much of the company's AI research and development work is completed by its Claude model. As of August 2026, Claude leads (AL4) approximately 26% of Anthropic's AI R&D work, up from less than 1% in February 2026, while over 90% of work is at AI collaboration level (AL3) or higher. The second metric tracks regulation and intervention of AI agents, noting about 30,000 agents execute tasks on internal platforms, with one operation per 47,000 intercepted by monitoring systems. The third metric discloses internal compute allocation: approximately 6% of compute for AI R&D was directed toward safety, rising to 12% in AI-led research. Anthropic CEO Dario Amodei had previously called for coordinated efforts to slow AI development, supported by OpenAI CEO Sam Altman and SpaceX. Anthropic stated these metrics aim to bridge the information gap between labs and the public, setting a benchmark for industry transparency.
Read sourceShow 1 older updatesHide older updates
Anthropic Reveals Its AI Claude Led 26% of R&D Tasks by August 2026, Up From Under 1%
On September 17, 2024, Anthropic published an article detailing internal metrics on AI R&D automation. The company disclosed that as of August 2026, its AI model Claude 'led' approximately 26% of Anthropic's AI research and development tasks, a sharp increase from less than 1% in February 2026. Anthropic introduced the 'Anthropic R&D Automation Index' based on Epoch AI's automation levels, noting that over 90% of tasks reached at least the 'AI collaboration' level (AL3), though no tasks were fully autonomous. The company also reported that about 30,000 AI agents were running simultaneously on its main internal platforms. Regarding compute allocation, Anthropic stated that 6% of all AI R&D computing power was used for AI safety work, and 12% of compute used for 'AI-driven AI R&D' was dedicated to safety. The company plans to continue publishing these metrics and submit to independent third-party verification. This disclosure is linked to CEO Dario Amodei's recent proposal to slow the pace of frontier AI development to allow more time for safety research.
Read source