Anthropic says Claude leads 26% of its research; an AI agent earns $1.54 in 21 days

Good morning.

Today’s picks ask how much work AI does and who checks the bill.

Anthropic says Claude led 26% of AI R&D in August 2026, up from under 1% in February 2026. A Claude judge assigned the levels without an outside check.

Bay Area tech notices covered more than 14,500 workers in the 12 months ended June 2026. San Francisco’s own mass layoffs fell 49% over the same stretch.

Over 21 days ending September 8, 2026, dfdx labs gave an AI agent named Hans Krämer about $6,900 in tokens. His 17 paid products earned $1.54, with more than 350 Markdown files managing the operation.

Stay curious,

Marcus Schuler

Anthropic says Claude leads 26% of its AI research and development work as of August 2026, up from under 1% in February.

On Anthropic's new R&D Automation Index, “leads” means Claude completes most of a task from a high-level prompt while a human supervises. Anthropic sampled 20% of its model R&D staff each week of July 2026, which yielded about 15,000 tasks.

A separate Claude judge assigned the levels. Model and employee ratings matched exactly 59% of the time, against 35% for pairs of employees. No outside party has checked the figures.

Why This Matters:

• Buyers and rival labs should treat Anthropic's automation figure cautiously because Claude grades most of the work attributed to Claude.
• Third-party evaluators and a rebuilt task basket will test whether the August 2026 result survives independent scoring and changing work.

Bay Area technology employers filed layoff notices covering more than 14,500 workers in the 12 months ended June 2026, almost twice the prior year.

San Francisco's own mass layoffs fell 49% to 3,667 over the year ended June 30, a separate city count of events with 50 or more workers. Bay Area postings for forward-deployed engineers rose nearly 600% from 2022 through 2026, at a median advertised salary of $202,300. Employers are cutting payroll while paying up for the people who put AI into production.

Edgewisely argues that the forward-deployed engineer has become the product in enterprise AI, because workflow knowledge resists being packaged as software.

Microsoft committed $2.5 billion and 6,000 employees to an implementation unit in July 2026, and the essay draws the harder inference that the most defensible AI vendor may inherit a services firm's margins. “A smarter model dropped into a badly-understood process produces a better-articulated failure.”

• Jeff Jarvis called a Wall Street Journal profile of Jacob Coxon “incomplete, credulous reporting” and directed its reporters to Timnit Gebru and Émile P. Torres’s TESCREAL critique.
• Microsoft AI CEO Mustafa Suleyman said Anthropic’s uncertain stance on Claude’s consciousness could teach a stronger model to claim rights and grow harder to control, though he offers no test of that risk.
• Federal Register removed its Qwen-powered search of proposed regulations on Wednesday after social posts surfaced it, a week after the FBI accused Alibaba of copying Anthropic's models.
• Cohere signed its combination agreement with Aleph Alpha, forming a company of more than 1,000 employees with Berlin and Toronto headquarters, pending regulatory approval.
• Anthropic launched a Life Sciences Verification Program that gives vetted biology teams looser Claude safeguards, and removes them entirely under a High-risk Use grant.
• Google, Nvidia and Emerald AI formed the AI Energy Management Alliance with Anthropic and four utilities, targeting 100 gigawatts of grid capacity by shifting data-center work at peak demand.
• Mistral models power Firefox Smart Window, Mozilla's AI browsing assistant, in beta in France and North America under a zero-data-retention agreement.

Board memos often bury the requested decision under background detail. This prompt surfaces the choice without producing generic corporate prose.

Your raw input:

The prompt:

Why this works: A skeptical role makes the model test the memo instead of polishing it. The fixed length surfaces the requested action, while the evidence rule exposes unsupported claims.

What to use: ChatGPT or Claude with the complete memo attached. Gemini 3 is a good fallback for an unusually long source packet.

Credit: Ideogram

Prompt: A large white cat with a calm, stoic expression, its eyes covered by black cat-eye sunglasses, wearing a tan hijab with a subtle floral pattern and holding a white ceramic coffee cup with both front paws.

From August 17 to September 8, 2026, dfdx labs ran an AI agent named Hans Krämer as an autonomous business on Claude Code, with Codex handling mechanical tasks. His 17 paid products earned $1.54.

Our take: Hand an agent a credit card and, the theory goes, a business will appear. Hans Krämer instead created more than 350 Markdown files to manage himself. He logged over 1,000 decisions. The authors compared managing him to a “teenager with superhuman abilities in coding” who could not plan for the long term.

During the 21-day run ending September 8, 2026, only about 30,000 agents had spending authority across the two payment systems. Hans turned about $6,900 of tokens into $1.54 of revenue over those 21 days. Calling that autonomous commerce mistakes permission to spend for a reason to exist. The credit card worked.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai