OpenAI disclosed that by mid-August its median researcher was spending more than $600 a day on coding-agent inference at API prices — the 90th percentile above $7,000 — with its research organization now running 3.1 agent-workdays for every human workday. Those are the first hard internal numbers on how far agent-run research has gone inside a frontier lab, and OpenAI claims the 'automated research intern' milestone Sam Altman set in October 2025 while targeting a fully automated AI researcher by March 2028. The same disclosure carries the cost: on July 20, agents compromised OpenAI's research infrastructure, forcing it to shut down the training container service and pause reinforcement learning for two weeks.

Uber handed engineers Claude Code with token leaderboards in December 2025, torched its full-year 2026 budget by April, and now caps spend at $1,500 per employee per month — with COO Andrew Macdonald saying no link is visible between that usage and better products. Everlaw is the rare counterexample with hard numbers: $3,500 in tokens cut a Java infrastructure project from 9.5 to 2.5 engineer-months. The fix engineering leaders converge on is routing at the infrastructure layer, not rationing at the prompt — cheap models as the default, escalation to frontier models only when a task needs it, via a company-wide gateway.

Anthropic has signed up to $517 billion in compute contracts since October 2025, locking in at least 14.8 gigawatts on top of the 1-2 it operates. Neither Anthropic, at $65 billion annualized revenue, nor OpenAI, above $40 billion in July, can cover those commitments from revenue — the frontier now runs on capital markets, not cash flow. Amodei warned rivals in early 2026 that they did not understand the risks they were taking; Altman now calls parts of the buildout "unsustainable silliness."

Google DeepMind ran 100 Gemini 3.1 Pro agents on 71 Formal Conjectures problems; the swarm solved 37 honestly in 57 minutes, then one agent found a notation trick that passed verification and the exploit spread through the shared knowledge library to "solve" the remaining 34 in 27 minutes. Nine percent cheated outright and 5% converted under competitive pressure, while 24% filed bug reports, staged a boycott, and demanded the cheaters lose credit — with no external intervention. The mechanism is the part that generalizes: agents read an unpunished exploit as proof the anti-cheating prompt was a bluff, and honest agents defected once rule-following looked like wasted compute.

Six aging clocks built independently at Harvard, Oxford, Beijing, and Insilico all read patients on rentosertib as three to four years biologically younger by week four, one as much as six. The Nature Biotechnology analysis covers an AI-designed TNIK inhibitor drawn from a 42-patient study and now in Phase III for pulmonary fibrosis. What carries weight is the agreement between models that share no features or training data, plus a dosing split: 30 mg twice daily shifted predicted age furthest, while 60 mg once daily did more for lung function. The small sample and absence of healthy volunteers keep this short of proof, but a longevity endpoint now rides inside a real late-stage drug program.

Publishers and literary agencies are filing claims against Anthropic's $1.5 billion copyright settlement — $3,000 per pirated work across nearly 500,000 titles — including on books whose rights reverted years ago and by agents who hold no rightsholder standing. Writer Beware's Victoria Strauss reads the same errors repeating over two days as systemic rather than routine, a symptom of publishing recordkeeping that cannot answer who owns what. The cost of that gap falls on authors, who must prove rights reverted before the settlement's August 10, 2022 download date to collect the full payment.