AI Is Now Out-Working the People Who Build It

Internal data from a frontier AI lab shows coding agents now produce roughly three workdays of research output per human workday, with token usage per researcher up 124-fold since December and typical daily agent costs exceeding $600. The lab has reached its 'automated research intern' milestone, while a separate autonomous research system placed 8th out of 4,000 teams in a reasoning-improvement competition. Meanwhile, a new poll finds 70% of Americans are more worried than excited about AI, and a small trial of an AI-designed drug showed potential anti-aging effects.

Inside the top AI labs, the machines have quietly started pulling more research hours than the humans who supervise them — and that's just the opening line of today's roundup. The most striking data point to surface today came from inside a frontier AI lab, where internal numbers show coding agents now logging roughly three full workdays of research output for every single workday a human researcher puts in. Token usage per researcher is up 124-fold since December, the typical researcher now burns north of $600 a day just running agents the heaviest users spend well over $7,000 , and about eight in ten researchers are juggling four or more agents at a time. The lab says it has now hit the "automated research intern" milestone its leadership publicly targeted for this month, with a fully automated AI researcher still on the roadmap for early 2028. It's not the only place this is happening: a separate autonomous research system built by another major AI player just placed 8th out of 4,000 teams in a competition measuring how well a system can teach a model to reason better — evidence, the company says, that AI can now improve AI at a level comparable to human experts. Put the two together and you get the clearest picture yet of why frontier labs guard their unreleased models so closely: the internal head start compounds fast, and it's now measurable in dollars and workdays, not just vibes. Medicine may have gotten a real, if early, proof point today too. A drug that AI designed from scratch — picking the target protein and drawing the molecule itself — was originally built to treat a lung-scarring disease, but new trial data shows something else: patients treated with it read as biologically younger across six independent "aging clock" models, with one analysis estimating an age drop of nearly three years. The sample is small, just 42 patients, and the dosing that showed the strongest anti-aging signal wasn't the one that best treated the underlying lung disease — suggesting two separate effects worth studying on their own. It's the kind of tangible medical result that AI industry leaders have argued could do more to shift public opinion than any amount of marketing, and it's an early test of that theory. Speaking of public opinion, a new national poll out today suggests it isn't shifting the way the industry might hope. Seventy percent of Americans say they're more worried than excited about AI, and — in a rarity for 2026 — that concern crosses party lines almost evenly. Usage keeps climbing just over half of adults now use AI regularly, up six points from last year even as trust stays low: only 18% say they trust AI-generated information most of the time. Opposition to local data centers is close to universal, seven in ten believe AI is already costing people jobs, and the vast majority think Washington's current AI rules don't go far enough. Asked which party they trust to handle AI policy well, the single largest group — 44% — said neither one. With data centers turning into a visible local flashpoint and no party earning the benefit of the doubt, this looks like it's shaping up into a real midterm issue rather than a background hum.