The 99% Solution

Buy the Nines

Anthropic shipped Claude Haiku 5.5 on Wednesday and everyone wrote the same sentence: the cheap model got 90% cheaper. True, and boring. The line nobody circled sits four rows down the page. For the first time on a Haiku-class model you can set the effort level, five of them, low to high. That isn't a price cut. That's a dial.

Here's what the dial does, and why thirty years of running money makes me stop on it. You have never been able to buy correctness by the yard. You hired a person, you got one price and one reliability you couldn't actually see, and you prayed. You could not walk in and say "give me my analyst, but four points more accurate, and tell me what the four points cost." Now you can. Point an agent at a job, tell it what "right" means, and it will quote you the price of being right 90% of the time, 95%, 99%. Correctness just became a line item. You buy the nines the job is worth and not one more.

That changes the whole fight. For two years we argued about which model is smartest. Wrong fight. Intelligence is now a commodity that deflates about 79x a year (we did that math in Two-Way Street), three credible open-weight models landed this week to keep it honest, and even Musk's own bot started routing jobs to Anthropic's Claude when the task called for it. The model is a dial tone. The game moved to operations: what do you need to be right about, how right, and what will you pay for the last nine.

But here's the part that bites the people who skip it. You can only buy a number you can measure, and you can only measure it with a ruler the agent doesn't get to hold. Tell the agent to grade its own homework and it optimizes to look right, not to be right. We watched Gemini do exactly that last week, benchmaxxing the test it knew was coming (No-Win Scenario). This week OpenAI dropped 372 math results and the mathematicians can't tell which are new, because the prompts were withheld. A result you can't check isn't a result. It's a rumor with a decimal point.

So the dial is real, and it's the best news an operator has had in years. It just comes with a bill nobody reads until it's structural: agents make work for agents, checkers checking checkers, and that bureaucracy bloats exactly where no one's watching the meter. The fix is the oldest one there is. Own the ruler, then buy the nines.

At COAI today: the full Signal/Noise, with the operator's three-step playbook and why an agent bureaucracy bloats wherever the decider isn't the payer, is live at getcoai.com.

Outsider Labs Advisory →

90%. That's how far Anthropic cut the list price of its cheapest model this week, down to a dime per million words in. The discount isn't the story. The control panel underneath it is: five effort levels, a dial that runs from cheap-and-good-enough to slow-and-sure. For the first time you can buy correctness the way you buy bandwidth. Pick the nines the job is worth, pay for those, and stop. You could never do that with a human, and most of your work doesn't need the last nine anyway.

For the investor, price the meter, not the model. Raw intelligence is deflating 79x a year and open weights keep the floor dropping, so margin on tokens compresses from here. The money accrues to the two things that don't commoditize: the harness that routes the work and the verifier that proves it. Underwrite the toll booth and the ruler, not the benchmark score.

For the business owner, build the ruler before you hire the robot. Don't deploy an agent until you've got a cheap, private check that defines "right" for the task, one the model can't see. Then make it quote you 90/95/99% and buy only the nines the job's value justifies. Most of your work isn't 99% work, and paying for nines you don't need is the new way to light money on fire.

For the parent, raise the judge, not the doer. The agent will produce the essay, the code, the analysis for a dime. The scarce, durable skill is the one the machine can't buy its way out of: knowing what "right" looks like and being able to check it. Teach taste, teach verification, teach the nerve to tell a confident machine it's wrong.

Thursday, October 8

— Harry and Anthony

Sources: