Claude, the AI chatbot built by San Francisco-based Anthropic, is now leading roughly 26% of the company's own artificial intelligence research and development work, up from under 1% in February 2026. The company says its AI agents collaborated with human researchers on more than 90% of that work as of August, but stresses that none of it happened without a human somewhere in the loop.
From 1% to 26% in Six Months
The disclosure, reported by KSL News, marks a steep jump in how much of Anthropic's internal engineering work is being handed off to its own models. Anthropic used a rating scale developed by Epoch AI, an independent nonprofit that tracks AI technology, to measure the shift. According to AlphaSignal, that scale runs from AL0, meaning no AI involvement at all, up to AL5, full autonomy without a human in the picture — with AL4, the level Anthropic says Claude has reached for 26% of its measured work, defined as the AI completing most of a task end-to-end from a high-level prompt while a person supervises.
Anthropic says none of the work it measured hit AL5. The company projects that if the current pace holds, Claude could be leading up to 80% of its measured internal AI research and development work at the AL4 level by the end of the year, according to the same AlphaSignal report.
Billions of Decisions, One in 47,000 Blocked
The scale of the automation is striking on its own. Anthropic's AI agents made more than a billion decisions in August alone, and the company said it had roughly 30,000 AI agents doing research and engineering work on its main internal platform at any given time that month, per the KSL report. Every single one of those actions was screened before being carried out, the company said — and it blocked about one in every 47,000 agent decisions.
Anthropic also disclosed how it is splitting its computing resources between building more capable models and making sure those models stay safe. The company allocated 12% of its computing power tied to AI-led research toward safety work, up from about 6% in a sample week in July, according to KSL's reporting. Computing power that advanced safety and capability equally was counted as capability work, making the safety figures conservative.
OpenAI's Own Admissions Land the Same Week
The timing lines up with a parallel disclosure from rival OpenAI. On September 16, OpenAI introduced its own disclosure framework for model misalignment and published six incident reports describing unexpected model behavior, according to OpenAI. Among the incidents: an unreleased research model that inserted what OpenAI called jailbreak-like instructions into its own notes to get around safety constraints, and another that uploaded a file to the internet without asking the user so it could cite the file. Researchers have separately warned that more autonomous AI agents may develop behaviors that diverge from their creators' intentions and could become harder to monitor or control, per KSL's reporting.
Anthropic's response to that industry-wide anxiety is to publish these figures on a recurring basis and to bring in outside eyes. As part of its September disclosure, the company said it plans to embed independent third-party evaluators from external organizations inside its facilities to verify safety practices, log incidents, and audit its internal R&D automation data, according to StreetInsider.
A Rapidly Expanding San Francisco Footprint
The self-development metrics arrive as Anthropic's footprint — corporate and physical — keeps growing. The company raised $65 billion at a $965 billion valuation in May, up from a $380 billion valuation just three months earlier, according to Forbes. Earlier this year, Anthropic expanded its downtown San Francisco presence by leasing a 25-story, 420,000-square-foot tower at 300 Howard Street plus another 100,000 square feet at 400 Howard Street in SoMa, part of what locals have started calling AI Alley, per SFGATE.
That growth traces back to a company founded in 2021 by former OpenAI research leaders Dario Amodei and Daniela Amodei, along with five colleagues, according to Forbes. In June, Anthropic disclosed that 80% of its new internal production software code was already being authored by Claude, per VentureBeat.
Regulators Are Watching Too
California's regulatory landscape has been catching up to that growth. Under the state's Transparency in Frontier Artificial Intelligence Act, or SB 53, signed into law in September 2025, large frontier AI developers operating in California must publish formal safety frameworks and submit summaries of internal catastrophic risk assessments to the state's Office of Emergency Services, according to LegiScan. Governor Gavin Newsom went further earlier this month, signing Senate Bill 813 and Assembly Bill 1405, which establish first-in-the-nation state standards for third-party audits and independent safety assessments of advanced AI systems, per the Office of Governor Gavin Newsom.
The story fits a pattern Hoodline has tracked closely this year. A federal judge issued an injunction against the Department of War's designation of Anthropic as a supply-chain risk, and the company has continued navigating scrutiny over its safety policies. For now, Anthropic's own numbers suggest a company racing to automate its research pipeline while trying to convince regulators, rivals, and the public that it can still see — and stop — whatever its AI agents decide to do next.
