Buy Wins, Not Players

Moneyball for Intelligence

Two pricing decisions are now pointed at each other like trains on one track. The WSJ reported last night that OpenAI is planning drastic price cuts — their word, drastic — anticipating a war for users with Anthropic. Anthropic, meanwhile, moves Fable 5 onto the meter in 12 days: $10 per million tokens in, $50 out. One lab is slashing the bill. The other is itemizing it. And in the middle sits the most lopsided number I've seen all year: Ramp's AI Index says the most AI-forward companies in America now spend $7,500 per employee per month on AI. The median US company spends $11.38.

There's a scene in Moneyball where Peter Brand finally says the quiet part to Billy Beane: "Your goal shouldn't be to buy players. Your goal should be to buy wins." The 2002 A's won as many games as the Yankees on about a third of the payroll, because Beane stopped paying for reputation and started paying for runs. Now look at what the AI market coughed up this spring: a cheap open model doing the labor with an expensive model supervising matched the expensive model working alone — at 39% of the cost. Harvey ran a hundred tasks on an open model for $84 that cost $954 on the frontier model, and the cheap one won. Same ratio as the A's. Nobody's drafting these box scores yet.

Here's the part the nerds fighting about benchmarks this week — and they are fighting, 12 million views' worth — keep missing. Good enough wins at every altitude. At the workflow level: Brian Armstrong says 80% of Coinbase's workloads land on models that are 99% cheaper, and only 20% need frontier IQ. At the market level: the models collapsing the price floor are open and largely Chinese — DeepSeek, GLM, Xiaomi pushing 1,000 tokens a second on eight commodity GPUs — which makes our export-control obsession with the last 10% look like guarding the wrong door. At the institutional level: the UAE just ordered half its federal government converted to agentic AI in two years, no debate, while we hold hearings. Tuition will be paid. Falling behind costs more.

The unclaimed prize is the routing layer — the thing that reads your workflows, sends 80% of the tokens somewhere nearly free, and reserves the meter for work that earns it. It barely exists as a product, because you can't route what you can't grade, and most shops haven't graded anything. That $7,500 is being spent at sticker price, no harness, no meter. The allocator who fixes that doesn't need a better model. Stop buying players. Start buying wins.

At COAI today: the full Signal/Noise — with the China deployment bet and the complete routing math — is live at getcoai.com.

Outsider Labs Advisory →

$7,500 vs. $11.38. That's the monthly AI spend per employee at America's most AI-forward companies versus the median US company, per Ramp's AI Index. Those aren't two budgets. They're two different decades running at the same time. And the $7,500 is being paid at full sticker — frontier models, no routing, no meter — right before a price war makes the same work radically cheaper. The gap isn't the scandal. The gap is the trade.

1. Open a token P&L before the meter starts. Anthropic re-tiers Fable 5 in 12 days; OpenAI is planning drastic cuts per the WSJ. You cannot negotiate, route, or even complain intelligently without knowing what your workflows cost. Today: pull last month's AI bills and tag every dollar to a workflow. One spreadsheet. An hour.

2. Run the 80/20 sort on your top ten AI workflows. Coinbase routes 80% of its work to models that are 99% cheaper; Harvey got better results for $84 than for $954. Today: mark each workflow "frontier required" or "good enough works," then actually downgrade one and measure the output. Most teams discover the expensive model was supervising email.

3. Pick one process and mandate it end-to-end — with a baseline. The UAE converts half its government in two years by directive; Block mandated AI and insiders saw little gain. The difference is the meter, not the will. Today: name one process, write down its current cost and cycle time, and set a 30-day clock. Will without measurement is how you become the cautionary tale.

Thursday, June 11

— Harry and Anthony

Sources: