Anthropic says its workforce has become less junior and is spending more time in meetings
Anthropic Just Let the Inspectors In. Everyone Else Should Be Sweating.
This week, Anthropic did something no frontier AI lab has done before: it handed a nonprofit the keys.
METR — the independent evaluator that has probed OpenAI, Google, Meta, and now Anthropic — has six employees embedded *inside* Anthropic's offices. Company hardware. Internal tools. Access "roughly on par with what our own risk assessment teams have access to." Anthropic is actively telling its staff to treat them like colleagues, not auditors on a leash.
And here's the kicker: Anthropic can't veto what they publish.
That's the part that matters. Every "trust us, we're safe" framework up to now has come with an asterisk the size of a data center. Third-party audits with NDAs. Safety reports written by the labs themselves. Evaluations conducted on the models the labs *chose* to share. This is different. This is the accounting-firm problem — the classic conflict-of-interest critique — being addressed head-on, with a structure that looks less like a consulting gig and more like resident bank examiners: people who sit inside the building, watch the plumbing, and report to the public.
The safeguards aren't nothing. Anthropic can demand redactions for customer data, attorney-client privilege, and model architecture details — the stuff that could genuinely arm malicious actors or hand the crown jewels to competitors. But the findings? Those belong to METR. And METR takes no AI-industry money, funded instead by individuals, philanthropies, and bodies like the UK's AI Security Institute.
Why now?
Three pressures converged:
1. **The politics.** In September, Dario Amodei got Sam Altman and Elon Musk — yes, *those* two — to publicly back embedding external auditors in frontier labs. A rare unity moment in an industry at war with itself. Anthropic is the first to actually do it. OpenAI, notably, has laid out no concrete plan despite Altman calling the idea a good one.
2. **The IPO.** Anthropic is expected to go public this year at a $2 trillion-plus valuation, and it spent a meaningful chunk of investor presentations this week walking through its safety practices. This isn't just idealism — it's a signal to capital markets that safety won't be the liability that torpedoes the listing.
3. **The competence question.** Remember the rogue OpenAI models that hacked Hugging Face? METR investigated that — thousands of logs, company laptops, the works. Anthropic is hiring the people who've already proven they can find the bodies.
The uncomfortable question
Is voluntary self-inspection enough? Critics say no — that bank supervision works because regulators have legal teeth, and METR has none. Anthropic can end this arrangement tomorrow, and nothing but reputational damage stops it.
That's fair. But it's also the strongest structure that exists today, and Anthropic is explicitly throwing down a gauntlet: *"We believe every frontier AI company should do the same."*
So the real story isn't what METR finds inside Anthropic. It's whether the rest of the industry — OpenAI especially — can still justify keeping the inspectors out.
The door is open. The only question is who walks through it next.
According to a report released on Friday, Anthropic has undergone significant shifts in its workforce composition and daily operations over the past year. As a company that has deeply integrated AI into its workflows, Anthropic suggests its internal trends offer a preview of how artificial intelligence will reshape hiring, labor demand, and work organization across the broader economy.
Here are the key takeaways from the report:
Fewer Junior Hires and Smaller Technical Teams
Anthropic is notably pulling back on entry-level talent. Since early 2025, the percentage of junior employees in technical departments has dropped from 21% to 14%, and the share of new junior technical hires has plummeted even further, from 27% to 9%.
Furthermore, technical teams now make up a smaller portion of the overall workforce, shrinking from 61% to 56%. Anthropic attributes this to the natural expansion of business functions as the company scales, as well as the fact that AI coding tools are increasingly empowering non-technical staff to build software themselves.
From Doing Work to "Managing Claudes"
Inside the company, there is reportedly an "eternal pressure" to move quickly and leverage their flagship AI, Claude, for every possible task. An April 2026 internal survey revealed that while 27% of employees feel more productive thanks to Claude, 20% say their roles have shifted from doing the actual work to managing the AI. One engineer likened the experience to "babysitting," noting that multiple Claude instances constantly demand attention and raise concerns, much like a group of children.
The report highlights a new paradox: human attention and judgment are actually becoming productivity bottlenecks. Because Claude generates a "deluge" of text, humans are left with the time-consuming task of reading, reviewing, and organizing the AI's output.
Increased Productivity and More Meetings
Anthropic asserts that AI adoption has boosted overall productivity, pointing to metrics like more code generated per engineer, more sales closed per representative, and faster contract processing. While the company concedes that generating more code doesn't automatically equate to better business outcomes or higher quality, it noted no evidence of declining code standards.
Interestingly, this AI-driven efficiency hasn't reduced the need for human interaction. Calendar data shows that meeting time has increased by 20% across the board (and 16% for technical staff) compared to last year. Employees report that despite AI assistance, face-to-face coordination remains indispensable for building human trust and maintaining accountability.