On September 8, ten thousand AI agents worked for 88 hours, ran up what OpenAI prices at about fifteen million dollars of computing, and proved that a still pond can be made to move infinitely fast. That same week, most of the world asked a chatbot to write a toast for a cousin’s wedding. Both groups used the same technology.

Five Stages of Noticing

The revolution has no front door, so people wander in through different rooms.

1. The Abstainer has heard of it and judges it a fad, like the Segway, for the same reason: nobody they know rides one.
2. The Toast Writer uses it for emails, wedding toasts, and a firm letter to the landlord, and is impressed and slightly embarrassed.
3. The Programmer describes software in English, receives software, and reads it with the wariness of a man meeting his own old code.
4. The Fleet Commander launches a dozen agents at bedtime and wakes to eleven finished tasks and one agent that has reorganised the kitchen cabinets.
5. The Swarm Operator launches ten thousand and tells nobody until the proof is in.

Most people stand at stage one or two. The news arrives from stage five.

What Is Actually Changing

The popular fear is the pink slip. The cause is plainer than the fear: the price of an hour of competent thinking is collapsing. Contract review, first-draft code, translation, data cleaning, literature searches. Each once cost a salaried hour. Each now costs pennies and seconds.

A job is a bundle of tasks held together by a salary. AI does not fire the bundle. It eats the toppings off the pizza, one task at a time, and leaves a person holding a smaller pizza and a larger responsibility. Entry-level work feels it first, because the first year of a career consists of exactly the tasks that are cheapest to hand over, and those tasks were the practice field.

Employers have noticed. Gartner forecasts that by 2027, 75% of hiring will test for AI proficiency, and that through 2026 half of organisations will require “AI-free” skills assessments, because leaning on chatbots dulls critical thinking. The candidate must prove they can drive the car and that they can still walk. Pilots live by the same rule: they train to fly with the autopilot and without it.

When production is cheap, checking becomes expensive. That sentence returns at the end.

The Fluid That Would Not Behave

Stir your coffee. The Navier-Stokes equations, written down in the 1800s, describe what happens next: how coffee swirls, how air bends over a wing, how blood moves through an artery. Engineers use them daily on supercomputers, because nobody can solve them by hand.

In 2000, the Clay Mathematics Institute attached a million dollars to one question about them. Can a smooth flow ever turn violent in finite time, with some speck of fluid reaching infinite speed? Mathematicians call that a blow-up, or a singularity. Your coffee says no, but a coffee cup is not a proof.

Twenty-six years passed. In September 2025, Google DeepMind and university teams used neural networks to find new families of singularities in related equations, the first sign that machines could scout this terrain. Then, on September 7, 2026, the mathematicians Alpöge and Buckmaster published a blow-up for the Euler equations, the cousin of Navier-Stokes that ignores a fluid’s stickiness, under a smooth outside push. The next morning OpenAI announced that 10,000 agents had carried the idea through to full Navier-Stokes in 88 hours, and that a second model had spent 17 more hours translating the proof into Lean, a language in which a computer checks every logical step. On September 11 the Clay Institute said the problem had apparently been settled.

Read the fine print. The proof concerns a fluid at rest that a smooth outside force drives to runaway speed. Clay’s wording admits such a force, so by its terms the prize question is answered. The unforced case, the pond left alone, which most mathematicians had pictured, stays open. And mathematicians are still checking the work, which is what mathematicians do to a miracle.

The physics adds a second footnote. The equations treat fluid as infinitely divisible. Zoom far enough and you meet molecules. One Brown University mathematician put the blow-up for air at vortices about 70 nanometres wide, a scale where individual molecules matter. The singularity is a confession by the equations, not a threat to your latte. The equations give out before the coffee does.

The bill: 2.7 million messages, about 130 billion output tokens, roughly $15 million at retail prices. That is fifteen times the prize money, the most human feature of the project: a heroic overspend on a trophy.

The Race

The labs compete on four fronts at once: capability, price, distribution, and trust.

OpenAI has the swarm and the momentum. Reports say it has clawed back market share from Anthropic this year, and OpenRouter data shows it topping Anthropic in weekly model spend for the first time in two and a half years. GPT-6 Astra and GPT-6.1 Sol sit close behind the leaders on a ranking published this week.

Anthropic sells trust to engineers. Its Claude Fable 5.1, Opus 5.5 and Sonnet 5.5 top that same ranking, its valuation reached $965 billion in May, and it launched Haiku 5.5 on October 7 at a 75% lower cost. The awkward part: The Information reports that Meta and Microsoft are pushing staff off Claude Code, which is what happens when your customer is also your competitor.

Google owns the chips and the data centers and arrived last week with Gemini 4 Argon. Analysts say it returns Google to the frontier conversation without making it the leader. Argon’s first assignment was tidying Google’s own garage: freeing hundreds of terabytes of memory in its data centers.

The open-weight crowd supplies cheaper models that anyone can download. Nvidia-backed Reflection AI is preparing a release, and Chinese labs such as DeepSeek, Moonshot and Alibaba already ship. Anthropic accuses several of them of distillation, training on another model’s answers, the machine equivalent of copying a classmate’s homework.

Meanwhile Crunchbase counts 195 AI startups bought by other AI companies in 2026, ten by OpenAI alone. An industry that eats itself this early is either mature or hungry.

What 2027 Holds

Gartner expects agents to mount the first real challenge to mainstream office software in thirty years, a $58 billion disruption, and expects fragmented AI regulation to cover half the world’s economies by 2027. The AI 2027 scenario, which foretold autonomous AI researchers and superintelligence, has 18 of its 53 predictions confirmed or running ahead of schedule as of June 2026. The dramatic ones are still due this winter and next year.

Three bets:

1. Swarms become a rented service. OpenAI first aimed at all six open Millennium Problems before narrowing to two. Prices keep falling, and Anthropic just cut 75% on its smallest model. The $15 million job costs a fraction by next autumn, and the other five Millennium Problems will meet the same treatment.
2. Agents talk to agents. Gartner expects 90% of business-to-business buying to run through AI agents by 2028. Your inbox fills with messages that one program wrote for another, and the sales cycle shrinks to a handshake between two strangers who are both software.
3. Checking becomes the profession. Lean verified the Navier-Stokes proof in hours; humans are still working out whether it says what everyone hopes. Auditors, referees, editors, reviewers. The scarce person of 2027 can say “this is correct” and be right.

Ten thousand agents solved a 26-year-old problem in under four days. A human, somewhere, is still checking it. Pay that human well.